Add Batch d1b948ea-b251-4189-b85d-fa6d52499e7c

ACL/2025/OmniAlign-V_ Towards Enhanced Alignment of MLLMs with Human Preference/04f08c85-e96b-4e1b-9f39-e98813dfc2d5_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3b1b44fb8e124940f0c86c8f878dda6e4dedd9822b9ccd2c8e7fcdba7a3bfa12
+size 175014

ACL/2025/OmniAlign-V_ Towards Enhanced Alignment of MLLMs with Human Preference/04f08c85-e96b-4e1b-9f39-e98813dfc2d5_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:41669d88e85ead6332ee6d2a4d9c0bb3fdd2195c81d49e33d4bea9c6f345660c
+size 225808

ACL/2025/OmniAlign-V_ Towards Enhanced Alignment of MLLMs with Human Preference/04f08c85-e96b-4e1b-9f39-e98813dfc2d5_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c7f42844b70f86c40edf50b37bb62845c4aaf59a914bcdd60ac417b68b9d73f2
+size 4381350

ACL/2025/OmniAlign-V_ Towards Enhanced Alignment of MLLMs with Human Preference/full.md

@@ -0,0 +1,945 @@
+# OmniAlign-V: Towards Enhanced Alignment of MLLMs with Human Preference
+
+Xiangyu Zhao $^{1,2*}$  Shengyuan Ding $^{2,3*}$  Zicheng Zhang $^{1,2}$
+
+Haian Huang² Maosong Cao² Weiyun Wang²,⁴ Jiaqi Wang² Xinyu Fang²,⁵
+
+Wenhai Wang $^{2}$  Guangtao Zhai $^{1,2}$  Haodong Duan $^{2\dagger}$  Hua Yang $^{1,2\dagger}$  Kai Chen $^{2\dagger}$
+
+Institute of Image Communication and Network Engineering, Shanghai Jiao Tong University<sup>1</sup>
+
+Shanghai AI Laboratory $^{2}$  Nanjing University $^{3}$  Fudan University $^{4}$  Zhejiang University $^{5}$
+
+![](images/081f3791503d2cfe8f24badc7f0c1f3e57041edb98d3cc8980eaa8629d68ba8e.jpg)  
+Figure 1: OmniAlign-V consists of curated images paired with open-ended, comprehensive question-answer pairs, significantly improving the alignment of MLLMs with human preferences. Additionally, we introduce MM-AlignBench, a human-annotated, high-quality benchmark specifically designed to evaluate the ability of MLLMs to align with human values.
+
+# Abstract
+
+Recent advancements in open-source multimodal large language models (MLLMs) have primarily focused on enhancing foundational capabilities, leaving a significant gap in human preference alignment. This paper introduces OmniAlign-V, a comprehensive dataset of 200K high-quality training samples featuring diverse images, complex questions, and varied response formats to improve MLLMs' alignment with human preferences. We also present MM-AlignBench, a human-annotated benchmark specifically designed to evaluate MLLMs' alignment with human values. Experimental results show that finetuning MLLMs with OmniAlign-V, using Supervised FineTuning (SFT) or Direct Preference Optimization (DPO), significantly enhances human preference alignment while maintaining or enhancing performance on standard VQA benchmarks, preserving their fundamental capabilities.
+
+# 1 Introduction
+
+With the rapid advancement of large language models (LLMs) (OpenAI, 2023a; Touvron et al., 2023), multi-modal large language models
+
+MLLMs (OpenAI, 2023b; Team et al., 2023) have also seen significant improvements. Most opensource MLLMs (Liu et al., 2023b; Chen et al., 2023b) are developed by connecting a vision encoder (Dosovitskiy, 2020) to a pretrained LLM, followed by vision instruction tuning. Existing vision instruction tuning datasets (Liu et al., 2023b; Chen et al., 2023a, 2024a) and multi-modal evaluation benchmarks (Liu et al., 2023d; Yue et al., 2023; Lu et al., 2023) primarily focus on assessing fundamental capabilities (object recognition, OCR, etc.) of MLLMs, while paying little attention to human preference alignment. Consequently, while open-source MLLMs achieve comparable performance to proprietary counterparts on objective metrics of these foundational skills, they display a significant gap in the alignment with human preferences, which detrimentally impacts user experience in multi-modal conversational interactions, as demonstrated in Appx. A.
+
+In this work, a preliminary investigation was conducted to quantitatively assess the degradation of human preference alignment in MLLMs(Sec. 3). Experimental results on text-only subjective evaluation benchmarks (Dubois et al., 2024; Li et al., 2024b) revealed that MLLMs exhibited a sub
+
+stantial performance drop compared to their corresponding LLMs. A plausible hypothesis suggests that MLLMs suffer from catastrophic forgetting during vision instruction tuning. Following this hypothesis, a straightforward solution would be incorporating both multi-modal and text-only SFT data (Xu et al., 2024; Cao et al., 2025; Ding et al., 2023) for joint learning. However, our experiments (Tab. 2) indicate that this approach not only failed to yield improvements in human preference alignment under multi-modal contexts but also demonstrated negative impacts on the foundational multi-modal skills of MLLMs. These observations suggest that enhancing human alignment in multi-modal scenarios necessitates the development of specialized multi-modal instruction tuning datasets.
+
+Existing multi-modal instruction tuning datasets predominantly focus on foundational capabilities, featuring simple language patterns and uniform response styles (Fig. 5 in appendix). This study argues that effective multi-modal datasets for enhancing human preference alignment should incorporate several critical characteristics: open-ended questions, broad topic coverage, diverse response formats (varying in length and style), and strict adherence to instructions. Based on these principles, we constructed OmniAlign-V. In terms of image sources, OmniAlign-V encompasses natural images and infographics such as posters and charts. Furthermore, a novel solution was developed to filter out semantically rich images from natural image collections. With respect to tasks, complex knowledge-based question answering, creative tasks, and reasoning tasks were designed for different image types. Each task category incorporates diverse subtasks, and state-of-the-art MLLMs were leveraged to obtain diverse, high-quality responses. Ultimately, OmniAlign-V comprises  $\sim 200\mathrm{K}$  multimodal instruction tuning samples, exhibiting a significantly different overall data distribution compared to existing multi-modal SFT datasets.
+
+Leveraging OmniAlign-V, comprehensive experiments were conducted to explore its full potential for enhancing human alignment in MLLMs. Integrating OmniAlign-V into the SFT stage of the LLaVA-NeXT (Liu et al., 2024a) structure with InternLM2.5-7B and Qwen2.5-32B yielded significant improvements in human preference alignment across various powerful LLM encoders, including InternLM2.5 (Cai et al., 2024) and Qwen2.5 (Bai et al., 2023a). Furthermore, on ground-truth-based
+
+VQA benchmarks like MMMU (Yue et al., 2023) or OCRBench (Liu et al., 2023e), MLLMs finetuned with OmniAlign-V displayed comparable or superior performance. Beyond its application in SFT, OmniAlign-V also demonstrated additional value when applied to Direct Preference Optimization (DPO) (Rafailov et al., 2024). Experimental results indicate that implementing OmniAlign-V for DPO yields further improvements in preference alignment, surpassing baseline models finetuned on the same dataset. After the SFT and DPO stage, our LLaVA-Next baseline with Qwen2.5-32B surpasses the state-of-the-art model Qwen2VL-72B (Chen et al., 2024b) finetuned on extensive proprietary datasets.
+
+Throughout the exploration, we also observed that existing multi-modal human preference benchmarks (Lu et al., 2024; Qian et al., 2024) lack diversity in image sources, contain repetitive questions, and lack clarity. To address these limitations, we introduce MM-AlignBench, a high-quality benchmark comprising 252 carefully curated samples with diverse image sources and meticulously crafted questions by human annotators, enabling comprehensive evaluation of MLLMs' alignment with human preferences. MM-AlignBench, along with other MLLM human preference benchmarks, was employed throughout this study for evaluation.
+
+The contributions of this study are as follows:
+
+1. An in-depth investigation into the degradation of MLLM human alignment, analyzing the impact of both text-only and multi-modal tuning data.  
+2. The introduction of OmniAlign-V, a comprehensive open-ended multi-modal SFT dataset, complemented by OmniAlign-V-DPO. Extensive experiments demonstrate the effectiveness of these datasets in improving human preference alignment.  
+3. The development of MM-AlignBench, a carefully curated benchmark comprising high-quality, human-annotated samples specifically designed to evaluate MLLMs' human preference alignment.
+
+# 2 Related Work
+
+Alignment of LLMs. Alignment, encompassing the ability to follow human instructions and provide helpful assistance (Liu et al., 2024b), has long been a critical aspect of LLMs. Recent works (Wang et al., 2023; Xu et al., 2024; Cao et al., 2025) focused on generating high-quality SFT training data to enhance the alignment of LLMs. Besides, recent benchmarks (Liu et al., 2023c; Li et al., 2023,
+
+2024b) have introduced open-ended and challenging questions to assess the alignment performance of LLMs. However, there is a notable lack of benchmarks designed to evaluate the alignment of MLLMs.
+
+Visual Question Answering Datasets. In the early stages, Visual Question Answering (VQA) datasets were primarily used for basic semantic alignment between images and text (Radford et al., 2021; Li et al., 2022; Zhao et al., 2024). These datasets were relatively simple in structure. Popular VQA datasets (Hudson and Manning, 2019; Antol et al., 2015) predominantly featured straightforward questions that elicited single-word or short-sentence answers. Similarly, visual classification and detection datasets (Lin et al., 2014; Schuhmann et al., 2021; Sharma et al., 2018) typically consisted of brief descriptions that focused on the main object within the image.
+
+Instruction Tuning data. With the rapid development of MLLMs (Chen et al., 2024b; Bai et al., 2023b), the instruction tuning data has gained more attention. LLaVA (Liu et al., 2023b) leveraged advanced LLMs to generate instruction-following formatted data from traditional VQA datasets. Recent works have further expanded this approach by employing MLLM to generate captions (Chen et al., 2023a), complex question-answer pairs (Chen et al., 2024a; Gu et al., 2024), OCR data (Carter, 2024) and math-related data (Shi et al., 2024). Other efforts (Tong et al., 2024; Li et al., 2024a) aggregates multiple publicly available document image datasets into unified resources. However, these datasets primarily focus on fundamental visual capabilities with short dialogues and factual questions, resulting in suboptimal alignment with human preferences.
+
+# 3 Human Alignment of MLLMs: The Preliminary Study
+
+When handling open-ended and flexible questions involving images, state-of-the-art open-source MLLMs - despite excelling in certain recognition tasks - exhibit significantly weaker alignment with human preferences compared to GPT-4o (Fig. 6 in appendix). We hypothesize that this decline in multi-modal preference alignment stems from a reduction in the language model's proficiency after the multi-modal SFT stage. To test this hypothesis, we evaluated state-of-the-art MLLMs on text-only human preference alignment bench
+
+<table><tr><td>Method</td><td>Type</td><td>AlignBench</td><td>AlpacaEval-V2</td><td>ArenaHard</td></tr><tr><td>InternLM2.5-7B</td><td>LLM</td><td>6.36</td><td>27.58</td><td>27.06</td></tr><tr><td>InternVL2-8B</td><td>MLLM</td><td>4.04(-36.4%)</td><td>3.35(-87.9%)</td><td>4.65(-82.8%)</td></tr><tr><td>LLaVA-Internlm</td><td>MLLM</td><td>4.66(-26.7%)</td><td>4.22(-84.7%)</td><td>4.93(-81.8%)</td></tr><tr><td>Qwen2-7B</td><td>LLM</td><td>6.02</td><td>24.47</td><td>32.84</td></tr><tr><td>Qwen2VL-7B</td><td>MLLM</td><td>4.92(-18.3%)</td><td>3.85(-83.4%)</td><td>6.46(-80.3%)</td></tr><tr><td>LLaMA3-8B</td><td>LLM</td><td>4.88</td><td>30.19</td><td>31.96</td></tr><tr><td>MiniCPM-V2.5</td><td>MLLM</td><td>3.84(-21.3%)</td><td>7.33(-75.7%)</td><td>8.05(-74.8%)</td></tr><tr><td>InternLM2-20B</td><td>LLM</td><td>5.49</td><td>43.35</td><td>33.75</td></tr><tr><td>InternVL2-26B</td><td>MLLM</td><td>4.39(-20.0%)</td><td>5.34(-87.7%)</td><td>11.25(-66.7%)</td></tr><tr><td>Hermes2-liama3-70b</td><td>LLM</td><td>5.72</td><td>46.09</td><td>57.40</td></tr><tr><td>InternVL2-76B</td><td>MLLM</td><td>4.33(-24.3%)</td><td>8.32(-81.9%)</td><td>16.17(-72.3%)</td></tr></table>
+
+Table 1: Language Alignment Benchmark Results. After multi-modal SFT, MLLMs demonstrate significant decline in human preference alignment compared to their corresponding LLMs.
+
+marks (Liu et al., 2023c; Li et al., 2023, 2024b). Additionally, we constructed a LLaVA (Liu et al., 2023b) baseline using InternLM-2.5-7B, fine-tuned on the LLaVA-Next-778k (Liu et al., 2023a) SFT dataset. As shown in Tab. 1, the LLM's ability to handle text-only open-ended questions degrades significantly after multi-modal SFT. This degradation may be due to (1) insufficient quantity or quality of text-only samples during multi-modal SFT, or (2) the overly simplistic style of multimodal fine-tuning data derived from traditional VQA datasets (Goyal et al., 2017; Hudson and Manning, 2019), often consisting of simple questions and short, factual answers.
+
+To assess the impact of the first factor on multimodal alignment, we examined the role of text-only data in multi-modal SFT. The LLaVA-Next-778K dataset includes approximately 40K text-only samples from ShareGPT (Chiang et al., 2023), which are relatively outdated and of lower quality. To improve alignment, we experimented with replacing these samples with better-quality alternatives and increasing their proportion in the multi-modal fine-tuning data. Specifically, we sampled 40K / 80K instances from two high-quality language SFT datasets: Magpie-Llama3.3 (Xu et al., 2024) and Condor (Cao et al., 2025), using these to replace the original language data in LLaVA-Next. We then evaluated models fine-tuned on these different data mixtures using both text-only benchmarks and multi-modal benchmarks, including WildVision (Lu et al., 2024) for human preference alignment and various benchmarks from OpenVLM Leaderboard (Duan et al., 2024) for assessing various fundamental multi-modal capabilities.*
+
+Results in Tab. 2 reveal several key findings. While MLLMs tuned with higher-quality text-only samples show significant improvements on text
+
+<table><tr><td rowspan="2">Model</td><td colspan="3">Language Benchmarks</td><td colspan="5">Multi-modal Benchmarks</td></tr><tr><td>AlignBench</td><td>AlpacaEval-V2</td><td>ArenaHard</td><td>WildVision</td><td>MM Vet</td><td>MMBench-V1.1</td><td>AI2D</td><td>OCRBench</td></tr><tr><td>\(LLaVA^I\)-\(LLaVANext778k\)</td><td>4.7</td><td>31.7</td><td>21.6</td><td>18.4/-55.1</td><td>41.2</td><td>73.7</td><td>74.2</td><td>39.7</td></tr><tr><td>\(LLaVA^I\)-\(LLaVANext_{mm}738k-Magpie40k\)</td><td>4.5 ↓</td><td>65.5 ↑</td><td>44.6 ↑</td><td>16.8/-58.9 ↓</td><td>37.7 ↓</td><td>73.1 ↓</td><td>73.4 ↓</td><td>38.7 ↓</td></tr><tr><td>\(LLaVA^I\)-\(LLaVANext_{mm}738k-Condor40k\)</td><td>5.6 ↑</td><td>72.9 ↑</td><td>55.7 ↑</td><td>16.8/-57.4 ↓</td><td>38.3 ↓</td><td>72.6 ↓</td><td>73.6 ↓</td><td>38.5 ↓</td></tr><tr><td>\(LLaVA^I\)-\(LLaVANext_{mm}738k-Condor80k\)</td><td>5.7 ↑</td><td>75.2 ↑</td><td>55.7 ↑</td><td>16.8/-57.0 ↓</td><td>38.3 ↓</td><td>72.6 ↓</td><td>74.0 ↓</td><td>37.6 ↓</td></tr></table>
+
+Table 2: Performance of Incorporating High-Quality Language Data.  $\mathsf{LLaVA}^I$  denotes LLaVA structure with InternLM2.5-7B as language model, and  $\mathsf{LLaVANext}_{mm}738\mathsf{k}$  refers to the multi-modal data in  $\mathsf{LLaVANext - }778\mathsf{K}$ . Integrating high-quality language data significantly improves alignment performance on language benchmarks. However, it leads to a decline in multi-modal alignment performance on benchmarks such as WildVision and MMVet. For AlpacaEval and ArenaHard, we present the winning rate against GPT-3.5.
+
+only alignment benchmarks, they unexpectedly demonstrate degraded performance on both multimodal alignment and common VQA benchmarks. This counter-intuitive phenomenon suggests that language alignment capability does not directly translate to multi-modal alignment. We therefore argue that high-quality multi-modal human-aligned data is crucial for improving MLLMs' human preference alignment in multi-modal contexts.
+
+# 4 OmniAlign-V
+
+Current MLLM instruction tuning datasets primarily focus on enhancing basic capabilities like perception, OCR, and mathematical reasoning. These datasets typically contain simple, brief question-answer pairs that inadequately capture human preferences and real-world interaction complexity, as shown in Fig. 5. We propose that multi-modal training data should also incorporate: (1) Open-ended, Diverse, and Creative Questions requiring interdisciplinary knowledge, (2) Comprehensive and Knowledge-rich Responses. To address these requirements, we develop a novel data synthesis pipeline to generate high-quality human-aligned multi-modal training data, resulting in the creation of OmniAlign-V.
+
+# 4.1 Task Taxonomy: A Big Picture
+
+Image content plays a crucial role in constructing multi-modal training data. To ensure comprehensive coverage, we classify images into two major categories: natural images and infographic images, as shown in Fig. 3(a). Our data synthesis pipeline first determines whether an input image belongs to natural images (captured from real-world scenes) or infographic images (human-created to convey information). Based on this classification, different vision-language tasks are assigned.
+
+For natural images, we define three primary tasks: Knowledge, Inferential, and Creation, each requiring diverse and complex question for
+
+mats with comprehensive, reasoned responses. These tasks enhance the model's ability to interpret real-world scenes effectively.
+
+For infographic images, given their diverse content, we identify four key types that elicit intricate and challenging questions: Arts, Charts, Diagrams, and Posters. These categories necessitate a deep understanding of human-designed visuals, including both abstract and detailed elements.
+
+# 4.2 Image Selection Strategy
+
+For natural images, rich semantic content leads to more comprehensive and insightful QA pairs. To enhance training data quality, we developed a pipeline to select semantically rich images from diverse sources, as shown in Fig. 2(a). This pipeline consists of two key steps.
+
+First, we use the Image Complexity (IC) recognition model IC9600 (Feng et al., 2022) to assign IC scores to images. Images with low semantic content — characterized by few objects or simple, uniform backgrounds — receive low IC scores and are then excluded. While IC9600 effectively filters out low-complexity images, high IC scores alone do not guarantee semantic richness. For instance, as shown in Fig. 7, an image filled with tents may have a high IC score but lacks sufficient semantic information for multi-modal training.
+
+To refine selection, we employ the Recognize Anything Model (Zhang et al., 2024) to identify objects within images, filtering out those with high complexity but minimal meaningful content. This two-step approach ensures our pipeline accurately selects images that are both complex and semantically rich. Experiment results have validated the effectiveness of our proposed image selection strategy, as shown in Appx. E.
+
+# 4.3 Data Generation Pipeline
+
+SFT QA Generation. In Fig. 2(b), we outline the generation process of SFT QA pairs. For vision-language tasks involving natural images
+
+![](images/09e2682da2c11438df13eb90cbae6b22900f53c49eb7351cc4305243684eabfb.jpg)
+
+![](images/5ccd88cac3c03a1a0f343b90e71330e16ed832dba516fe02383a0cf365b2639f.jpg)
+
+![](images/33fbf9219053d301e9cf803843dbead91a91289df5a92672ea4dabab8d88ccb7.jpg)
+
+![](images/a662bceea221ef26d138856e872d82fe4ce8920f13a8af8390facdd207c1b963.jpg)  
+Figure 2: Overall pipeline of OmniAlign-V. By utilizing an image filter and employing a customized pipeline for distinct tasks, we curate semantically rich images paired with high-quality open-ended question-answer sets. Post-refinement further enhances both the variety and quality of our dataset.
+
+(knowledge, creative, inferential), we first apply our image selection strategy to filter semantically rich images from CC-3M(Sharma et al., 2018), Flickr30k (Plummer et al., 2015), and GQA (Hudson and Manning, 2019). For infographic image tasks, images are collected from various existing sources (Masry et al., 2022; Li et al., 2018; Kembhavi et al., 2016). More details are provided in Appx. B.
+
+Knowledge & Inferential Tasks. In preliminary experiments, we found that GPT-4o can generate diverse, content-relevant questions for the two tasks when provided with well-designed few-shot prompts. Consequently, we carefully designed a single prompt for each task category, incorporating comprehensive instructions and selected few-shot examples. These prompts were then used with GPT-4o across diverse images to generate QA pairs.
+
+Creative Tasks. We noticed that a single prompt cannot generate sufficiently diverse, content-relevant creative questions. Therefore, we developed a more sophisticated pipeline inspired by Condor (Cao et al., 2025). First, we created a set of seed creative questions:
+
+$$
+\mathcal {Q} _ {s} = \left\{Q _ {1}, Q _ {2}, \dots Q _ {N} \right\} \tag {1}
+$$
+
+where each seed question corresponds to a distinct creative task. Since directly using all seed questions as few-shot examples leads to repetition
+
+and lack of diversity, we employ a light-weight MLLM (Luo et al., 2024) to generate detailed captions  $C$  for each image. An LLM  $\mathcal{M}$  then selects a relevant subset of seed questions according to the caption:
+
+$$
+\mathcal {Q} _ {s} ^ {\prime} = \mathcal {M} (C, \mathcal {Q} _ {s}), \| Q _ {s} ^ {\prime} \| \ll \| Q _ {s} \| \tag {2}
+$$
+
+Finally, we randomly select three question types from  $\mathcal{Q}_s'$  as few-shot examples for GPT-4o, preserving both quality and diversity of synthesized data.
+
+Infographic Tasks. For infographic image tasks (Charts, Diagrams, etc.), questions and answers are closely tied to specific image content. Unlike natural images, these visuals convey information primarily through text, colors, lines, and symbolic elements, making evaluation based on image complexity or object categories unsuitable. Therefore, instead of applying image selection strategies, we carefully select image sources containing rich, detailed information. We then design specialized prompts for GPT-4o to generate questions that require comprehensive background knowledge understanding. The difference is shown below:
+
+InfographicVQA: What is the respiratory disease death rate for individuals aged  $70+$ ?
+
+OmniAlign-V: How does the respiratory disease death rate for individuals aged  $70+$  compare to the other age groups, and what might this suggest?
+
+![](images/ccff641fa8eb0c88734f547e5508bacb0c20e1ea0eda6c9c660ec352987776c8.jpg)  
+(a) Distribution of tasks.
+
+![](images/9e3a3a4c853f013ca600392d3f33eb0da0e10710e92d28115f2b53326214ab6a.jpg)  
+(b) Comparison with other datasets.
+
+Post Refinement. To further improve the quality of the synthesized data, we implemented a series of post-processing methods for refinement, as illustrated in Fig. 2(c).
+
+Instruction Augmented Knowledge QAs. Instruction following is a crucial capability significantly impacting human preference. To enhance this, we incorporate complex instructions and restrictions into our knowledge QA pairs and reformulate responses accordingly. As shown in Fig. 2(c1), for each knowledge QA, we use a powerful LLM to select an appropriate instruction type that can be integrated into the question without depending on visual content. The instruction is then incorporated into the existing question, and an LLM adjusts the corresponding answer to ensure alignment with both the modified question and original context, resulting in instruction-augmented knowledge QAs.
+
+Enriched Inferential QAs. For many inferential QAs, the answers lack sufficient detail to fully explain underlying logic and background knowledge. To address this, we employ a knowledge-rich LLM to enrich responses with detailed explanations, relevant background information, and logical reasoning (Fig. 2(c2)). This refinement enhances alignment with user preferences.
+
+Quality Improved Infographic QAs. For infographic tasks, particularly those involving Chart data, we observed that even SOTA MLLMs struggle with complex charts and detailed questions. While GPT-4o excels at explaining background knowledge but often produces inaccurate OCR results, SOTA open-source MLLMs (Chen et al., 2024b; Bai et al., 2023b) show superior OCR accuracy but lack detailed explanations. Therefore, we developed a refinement pipeline to generate responses combining rich background knowledge
+
+![](images/2eaaa931a184d2b05fdf8c13067463ccb95c05c49c600763d896a7c323edbf7f.jpg)  
+Q: What are the potential challenges and necessary preparations for vehicles traveling in the environment depicted in the image?
+
+![](images/970e974d4211d75adc7efaa670bc25f398a2e439f0df93952b8a8f5cad2071c8.jpg)  
+Figure 3: Data distribution of OmniAlign-V. Our dataset includes a diverse range of tasks, characterized by a more balanced distribution of answer lengths compared to those observed in ALLaVA and ShareGPT4V.  
+Q: What does the data suggest about the historical or cultural influences on foreign policy preferences in East and West Germany?  
+Figure 4: Samples in MM-AlignBench.
+
+![](images/fa7b2ad1956754b6cb117d306d71e2ed8038bbd52253c24c2daa28edbb8b94fd.jpg)  
+Q: Imagine stepping into the shoes of the character in the image. Deliver an inspirational pep talk to a team, drawing on the energy and intensity of the moment.
+
+![](images/026c1e743aa89d931dbd6ba43bd5b06ea4fd8336ee564371e92e4840afe8a2b0.jpg)  
+Q: What role does the Metamorphic rock play in this diagram?
+
+and accurate OCR results. We filter out questions where GPT-4o and SOTA open-source MLLM responses show significant discrepancies in trends. For remaining questions, the responses from different MLLMs are merged to produce a final response that is both precise and comprehensive. The merged answers are further reviewed by human experts to ensure their quality and consistency. More details are demonstrated in Appx. C.
+
+OmniAlign-V comprises 39K Knowledge QAs, 37K Inferential QAs, 10K Creative QAs, 38K Instruction-Following (Knowledge) QAs, and 44K Infographic QAs (2K Art, 8K Diagram, 11K Chart, 23K Poster). Additionally, we prompt GPT-4o to generate 35K QAs focusing on image details, resulting in a total of 205K high-quality SFT training samples. Examples are shown in Figs. 9 to 12.
+
+DPO Data Generation. OmniAlign-V's high-quality, human-aligned QA pairs can serve as positive samples for DPO training. Inspired by Reject Sampling (Casella et al., 2004), we generate negative samples by prompting a LLaVA-Next baseline (generator  $G$ ) trained on LLaVA-Next-778k. For each question  $Q_{i}$ , the generator produces  $N$  responses with high temperature,  $\mathbf{R} = \{r_1^i, r_2^i, \ldots, r_N^i\}$ . An LLM Judger  $J$  then evaluates these responses to select the one that most deviates from the original question's intent and context as the negative sample  $r_{Neg}^i$ , ensuring clear contrast between positive and negative samples.
+
+# 5 MM-AlignBench
+
+Current benchmarks for assessing multi-modal alignment capabilities are limited. While WildVision (Lu et al., 2024) aims to evaluate human preferences in real-world interactions, it employs repetitive and simplistic question formats that inadequately assess response quality, as shown in Fig. 8.
+
+<table><tr><td>Model</td><td>Data</td><td>LLM</td><td>MM-AlignBench</td><td>WildVision</td><td>MIA-Bench</td><td>MM Vet</td><td>MMMU</td><td>MMBenchV1.1</td><td>AI2D</td><td>OCRBench</td></tr><tr><td>LLaVA</td><td>LLaVANext-778k</td><td>InternLM2.5-7B</td><td>3.6 / -82.1</td><td>18.4 / -55.1</td><td>75.4</td><td>41.2</td><td>42.6</td><td>73.6</td><td>74.1</td><td>39.7</td></tr><tr><td>LLaVA</td><td>OmniAlign-Vmix</td><td>InternLM2.5-7B</td><td>50.0 / +3.8</td><td>28.2 / -34.6</td><td>85.4</td><td>43.5</td><td>43.3</td><td>73.7</td><td>74.7</td><td>41.3</td></tr><tr><td></td><td></td><td></td><td>+46.4 / 85.9</td><td>+9.8 / 20.5</td><td>+10.0</td><td>+2.3</td><td>+0.7</td><td>+0.1</td><td>+0.6</td><td>+1.6</td></tr><tr><td>LLvANext</td><td>LLvANext-778k</td><td>InternLM2.5-7B</td><td>20.6 / -42.7</td><td>23.4 / -45.0</td><td>76.9</td><td>41.8</td><td>44.1</td><td>75.1</td><td>74.7</td><td>56.2</td></tr><tr><td>LLvANext</td><td>OmniAlign-Vmix</td><td>InternLM2.5-7B</td><td>57.1 / +11.1</td><td>29.6 / -31.3</td><td>86.7</td><td>47.7</td><td>46.8</td><td>74.9</td><td>77.5</td><td>58.9</td></tr><tr><td></td><td></td><td></td><td>+36.5 / 53.8</td><td>+6.2 / 13.7</td><td>+9.8</td><td>+5.9</td><td>+2.7</td><td>-0.2</td><td>+2.8</td><td>+2.7</td></tr><tr><td>LLvANext</td><td>LLvANext-778k</td><td>Qwen2.5-32B</td><td>26.6 / -29.0</td><td>25.2 / -41.3</td><td>86.0</td><td>47.7</td><td>55.2</td><td>79.3</td><td>79.6</td><td>55.9</td></tr><tr><td>LLvANext</td><td>OmniAlign-Vmix</td><td>Qwen2.5-32B</td><td>62.3 / +19.4</td><td>40.2 / -14.9</td><td>89.6</td><td>56.9</td><td>60.7</td><td>80.6</td><td>81.7</td><td>55.9</td></tr><tr><td></td><td></td><td></td><td>+35.7 / 48.4</td><td>+15.0/26.4</td><td>+3.6</td><td>+9.2</td><td>+5.5</td><td>+1.3</td><td>+2.1</td><td>+0.0</td></tr></table>
+
+Table 3: Evaluation Results of Integrating OmniAlign-V into the SFT stage. By integrating OmniAlign-V, the multi-modal preference alignment of MLLMs significantly improved. Additionally, we also observe comparable or better performance on common VQA benchmarks. In 'Model' column, LLaVA and LLaVANeXT denote the model structure and training strategy. For MM-AlignBench and WildVision, notation  $A/B$  denotes Winning Rate/Rewards.  
+
+<table><tr><td rowspan="2">Model</td><td rowspan="2">Type</td><td>AlpacaEvalv2</td><td>ArenaHard</td></tr><tr><td colspan="2">v.s. GPT-3.5/GPT-4</td></tr><tr><td>LLaVANextI</td><td>MLLM</td><td>29.8 / 3.8</td><td>21.4 / 4.93</td></tr><tr><td>LLaVANextI-OA</td><td>MLLM</td><td>50.1 / 7.8</td><td>30.4 / 9.33</td></tr><tr><td>InternLM2.5-7B</td><td>LLM</td><td>78.3 / 26.2</td><td>47.5 / 19.1</td></tr><tr><td>LLaVANextQ</td><td>MLLM</td><td>50.6 / 7.0</td><td>54.5 / 18.0</td></tr><tr><td>LLaVANextQ-OA</td><td>MLLM</td><td>77.6 / 18.0</td><td>87.2 / 58.1</td></tr><tr><td>Qwen2.5-32B</td><td>LLM</td><td>92.1 / 37.1</td><td>94.8 / 75.9</td></tr></table>
+
+Table 4: Performance on text-only alignment benchmarks. OA denotes models trained with OmniAlign-V. We present the winning rate against GPT-3.5 and GPT-4 (original setting). OmniAlign-V can also enhance MLLM's performance on text-only alignment benchmarks.
+
+To address this, we developed MM-AlignBench, featuring human-curated images and questions for more nuanced evaluation.
+
+We selected high-quality images from SAM1B (Kirillov et al., 2023), CC-3M-Test (Sharma et al., 2018), AI2D (Kembhavi et al., 2016), ChartQA (Masry et al., 2022), and InfographicVQA (Mathew et al., 2022). For natural images, we applied our image selection strategy from Sec. 4.2 to identify 2,000 semantically rich images and combine them with 1,000 carefully selected infographics images. GPT-4o was then used to generate diverse questions for these images. Human experts reviewed and refined the image-question pairs, filtering out low-quality, repetitive, or contextually weak samples. This process resulted in 252 high-quality question-image pairs, featuring diverse question types and semantically rich images. Several examples are shown in Fig. 4.
+
+For evaluation, we follow WildVision's approach, using GPT-4o as the judge model to compare model responses with reference responses generated by Claude3V-Sonnet (Anthropic, 2024), reporting winning rates and reward scores.
+
+# 6 Evaluation Results
+
+# 6.1 SFT with OmniAlign-V
+
+We conduct extensive experiments to demonstrate OmniAlign-V's effectiveness. We combine OmniAlign-V with LLaVA-Next-778k (excluding text-only samples), creating OmniAlign- $\mathbf{V}_{\text {mix }}$  with 946K training samples. We evaluate various MLLMs tuned on OmniAlign-V against their counterparts tuned on LLaVA-Next-778k.
+
+Our evaluation spans multiple multi-modal benchmarks, including standard VQA benchmarks (Yu et al., 2023; Liu et al., 2024c, 2023e; Yue et al., 2023; Kembhavi et al., 2016) and human-preference alignment benchmarks: MMAlignBench, WildVision (Lu et al., 2024), and MIA-Bench (Qian et al., 2024). Results in Tab. 3 show that OmniAlign-V significantly improves human alignment across all benchmarks. Moreover, our training data improves general multi-modal capabilities, particularly on MMVet and MMMU, demonstrating a trend distinct from text-only data.
+
+Notably, despite excluding language samples from training data, models maintain stronger language alignment than those trained on LLaVA-Next-778k, as shown in Table 4. This suggests that while high-quality language data alone may not significantly impact multi-modal capabilities, enhancing multi-modal data quality can improve both language and multi-modal performance, highlighting the crucial role of high-quality, human-aligned multi-modal training data.
+
+# 6.2 DPO with OmniAlign-V-DPO
+
+We conduct DPO post-training on three models: LLaVA-Next trained with LLaVA-Next-778k, LLaVA-Next trained with OmniAlign- $V_{mix}$ , and InternVL2-8B. Results in Tab. 5 show that DPO tuning significantly improves performance on real-world questions (WildVision) across all models.
+
+<table><tr><td>Model</td><td>Stage</td><td>MM-AlignBench</td><td>WildVision</td></tr><tr><td>LLaVANextI</td><td>SFT</td><td>9.5 / -69.2</td><td>30.4 / -34.2</td></tr><tr><td>LLaVANextI</td><td>SFT+DPO</td><td>11.1 / -64.5</td><td>35.5 / -23.4</td></tr><tr><td>LLaVANextI-OA</td><td>SFT</td><td>57.1 / +11.1</td><td>29.6 / -31.3</td></tr><tr><td>LLaVANextI-OA</td><td>SFT+DPO</td><td>64.3 / +22.4</td><td>41.8 / -10.1</td></tr><tr><td>InternVL2-8B</td><td>SFT</td><td>31.4 / -21.8</td><td>48.6 / +1.4</td></tr><tr><td>InternVL2-8B</td><td>SFT+DPO</td><td>64.7 / +19.4</td><td>51.4 / +1.9</td></tr></table>
+
+Table 5: Performance of applying DPO with OmniAlign-V-DPO. For models finetuned with human-aligned data, by employing DPO training, the model's alignment with human perference further improved.  
+
+<table><tr><td>Model</td><td>Win Rate ↑</td><td>Reward↑</td><td>B+</td><td>B</td><td>T</td><td>W</td><td>W+</td></tr><tr><td>Claude3.5V-Sonnet</td><td>84.9</td><td>+51.4</td><td>70</td><td>144</td><td>13</td><td>25</td><td>0</td></tr><tr><td>GPT-4o</td><td>81.3</td><td>+49.0</td><td>81</td><td>124</td><td>12</td><td>31</td><td>4</td></tr><tr><td>GPT-4V</td><td>82.5</td><td>+46.0</td><td>57</td><td>151</td><td>12</td><td>31</td><td>1</td></tr><tr><td>GeminiFlash1.5-002</td><td>77.0</td><td>+39.1</td><td>56</td><td>138</td><td>14</td><td>35</td><td>9</td></tr><tr><td>LLaVANext-OA-32B-DPO</td><td>74.2</td><td>+36.9</td><td>49</td><td>138</td><td>20</td><td>40</td><td>5</td></tr><tr><td>Qwen2VL-72B</td><td>61.5</td><td>+21.6</td><td>43</td><td>112</td><td>15</td><td>75</td><td>7</td></tr><tr><td>LLaVANext-OA-32B</td><td>62.3</td><td>+19.4</td><td>31</td><td>126</td><td>19</td><td>62</td><td>14</td></tr><tr><td>Claude-3V-Sonnet</td><td>50</td><td>0</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>Qwen2VL-7B</td><td>44.4</td><td>-5.8</td><td>28</td><td>84</td><td>5</td><td>101</td><td>34</td></tr><tr><td>InternVL2-72B</td><td>44.4</td><td>-6.9</td><td>19</td><td>93</td><td>8</td><td>98</td><td>34</td></tr><tr><td>InternVL2-8B-MPO</td><td>40.1</td><td>-10.9</td><td>26</td><td>75</td><td>10</td><td>100</td><td>41</td></tr><tr><td>InternVL2-8B</td><td>31.3</td><td>-21.8</td><td>18</td><td>61</td><td>15</td><td>109</td><td>49</td></tr><tr><td>LLaMA3.2-Vision-11B</td><td>27.8</td><td>-33.7</td><td>18</td><td>52</td><td>4</td><td>98</td><td>80</td></tr><tr><td>LLaVANext-Qwen32B</td><td>26.6</td><td>-29.0</td><td>16</td><td>51</td><td>10</td><td>121</td><td>54</td></tr><tr><td>LLaVA-OneVision-7B</td><td>23.8</td><td>-46.2</td><td>14</td><td>46</td><td>1</td><td>75</td><td>116</td></tr><tr><td>MiniCPM-V-2.5</td><td>12.7</td><td>-53.0</td><td>9</td><td>23</td><td>8</td><td>116</td><td>96</td></tr><tr><td>Xcomposer2.5-7B</td><td>7.5</td><td>-74.0</td><td>5</td><td>14</td><td>3</td><td>63</td><td>167</td></tr><tr><td>Idefics3-8B</td><td>2.7</td><td>-92.3</td><td>3</td><td>4</td><td>0</td><td>15</td><td>230</td></tr></table>
+
+While the baseline trained solely on LLaVA-Next-778k shows minimal improvement on MMAlignBench, models incorporating OmniAlign-V during SFT demonstrate substantial gains after DPO. Similarly, InternVL2-8B, a state-of-the-art MLLM partially trained on proprietary multimodal corpora, shows significant improvement on MM-AlignBench post-DPO. This indicates that if a model has been trained on data aligned with human preferences, such as open-ended or long-context data during SFT phase, the subsequent DPO training on high-quality human-aligned data can significantly activate the model's ability, leading to a considerable improvement in alignment performance. In contrast, if the model has not been exposed to such alignment-focused datasets during SFT, training with open-ended data alone via DPO will not significantly improve its capabilities of alignment. These findings demonstrate the value of OmniAlign-V in both SFT and DPO stages for enhancing human preference alignment.
+
+Table 6: Performance of existing MLLMs on MM-AlignBench. B+/B/T/W/W+ denotes MuchBetter/Better/Tie/Worse/MuchWorse. Our LLaVA-NextOmniAlign(OA)-32B-DPO, trained with OmniAlign-V and applied DPO with OmniAlign-V-DPO, demonstrates outstanding performance, surpassing a wide range of strong MLLMs, even Qwen2VL-72B.  
+
+<table><tr><td>Model</td><td>MM-AlignBench</td><td>WildVision</td><td>MM Vet</td></tr><tr><td>LLaVANext-778k</td><td>20.6 / -42.7</td><td>23.4 / -45.0</td><td>41.7</td></tr><tr><td>+Know / Infer / Detail</td><td>23.4 / -42.1</td><td>23.2 / -43.2</td><td>42.8</td></tr><tr><td>+Instruction Following</td><td>36.5 / -17.3</td><td>26.2 / -37.5</td><td>44.6</td></tr><tr><td>+Creation</td><td>43.7 / -5.0</td><td>26.6 / -37.7</td><td>44.4</td></tr><tr><td>+Chart / Diagram / Poster</td><td>57.1 / +11.1</td><td>29.6 / -31.3</td><td>47.7</td></tr></table>
+
+Table 7: Ablation of Subsets in OmniAlign-V. By progressively incorporating tasks within OmniAlign-V, the model's alignment performance gradually improves.
+
+# 6.3 MM-AlignBench
+
+We evaluate various state-of-the-art MLLMs (OpenAI, 2023a; Team et al., 2023; Anthropic, 2024; Bai et al., 2023b; Chen et al., 2024b; Li et al., 2024a; OpenBMB, 2024; Dong et al., 2024; Meta, 2024; Laurencon et al., 2024; Wang et al., 2024) on MM-Alignbench, with results shown in Tab. 6. Closed-source models like GPT, Claude, and Gemini demonstrate strong alignment in responding to open-ended questions. In contrast, while Qwen2-VL and InternVL2 excel at common VQA tasks, they show relatively lower human preference alignment. This highlights the importance of prioritizing MLLM alignment for improved daily human interactions. Our LLaVA-OA-32B, trained with OmniAlign-V, achieves exceptional performance, outperforming numerous strong MLLMs and nearly matching Qwen2VL-72B. After applying DPO with OmniAlign-V-DPO, LLaVA-OA-32B-DPO achieves winning rate of 72.6 with an average reward of +33.5, surpassing the performance of Qwen2VL-72B. These results highlight the high quality and effectiveness of the OmniAlign-V dataset in improving model alignment with human preferences.
+
+# 6.4 Ablation Study
+
+We conduct an ablation study to evaluate each subset of OmniAlign-V, reporting results on MMAlignbench, WildVision, and MMVet in Tab. 7. Performance improves progressively as different tasks from OmniAlign-V are incorporated. Notably, Instruction-Following data significantly enhances performance across all three benchmarks, demonstrating its crucial role. The creation data subset uniquely improves performance on MMAlignbench but not on WildVision and MMVet. This discrepancy can be attributed to the absence of multi-modal creative question types in these two benchmarks, suggesting their incompleteness in capturing full spectrum of alignment challenges.
+
+# 7 Conclusion
+
+In this paper, we introduce OmniAlign-V, a dataset designed to enhance the alignment of MLLMs with human preferences, as well as MM-AlignBench, a high-quality, specific-purpose benchmark for evaluating such alignment. We investigate the impact of both language and multi-modal training data, emphasizing the critical role of multi-modal open-ended training data. By incorporating OmniAlign-V into SFT and DPO stages, we achieve significant improvements in the alignment of MLLMs.
+
+# 8 Limitation
+
+Although the OmniAlign-V pipeline can be easily scaled to support larger datasets, the scale of the dataset used in this paper may be insufficient for large-scale training due to cost limitations. Deeper exploration into the alignment of MLLMs is still needed to address these limitations and further advance the field.
+
+# 9 Acknowledgements
+
+This research was partly supported by grants of National Natural Science Foundation of China (NSFC, Grant No. 62171281), Science and Technology Commission of Shanghai Municipality (STCSM, Grant Nos. 20DZ1200203, 2021SHZDZX0102), National Key R&D Program of China (No.2022ZD0161600), the Shanghai Postdoctoral Excellence Program (No.2023023), China Postdoctoral Science Fund (No.2024M751559), and Shanghai Artificial intelligence Laboratory.
+
+# References
+
+Anthropic. 2024. The claude 3 model family: Opus, sonnet, haiku.  
+Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. 2015. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 2425-2433.  
+Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Junyang Lin, Runji Lin, Dayiheng Liu, Gao Liu, Chengqiang Lu, Keming Lu, Jianxin Ma, Rui Men, Xingzhang Ren, Xuancheng Ren, Chuanqi Tan, Sinan Tan, Jianhong Tu, Peng Wang, Shijie Wang, Wei Wang, Shengguang Wu, Benfeng Xu, Jin Xu, An Yang, Hao Yang, Jian Yang, Shusheng Yang, Yang Yao, Bowen Yu, Hongyi Yuan, Zheng Yuan, Jianwei Zhang, Xingxuan Zhang, Yichang Zhang, Zhenru Zhang, Chang
+
+Zhou, Jingren Zhou, Xiaohuan Zhou, and Tianhang Zhu. 2023a. Qwen technical report. arXiv preprint arXiv:2309.16609.  
+Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. 2023b. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966.  
+Zheng Cai, Maosong Cao, Haojiong Chen, Kai Chen, Keyu Chen, Xin Chen, Xun Chen, Zehui Chen, Zhi Chen, Pei Chu, Xiaoyi Dong, Haodong Duan, Qi Fan, Zhaoye Fei, Yang Gao, Jiaye Ge, Chenya Gu, Yuzhe Gu, Tao Gui, Aijia Guo, Qipeng Guo, Conghui He, Yingfan Hu, Ting Huang, Tao Jiang, Penglong Jiao, Zhenjiang Jin, Zhikai Lei, Jiaxing Li, Jingwen Li, Linyang Li, Shuaibin Li, Wei Li, Yin ing Li, Hongwei Liu, Jiangning Liu, Jiawei Hong, Kaiwen Liu, Kuikun Liu, Xiaoran Liu, Chengqi Lv, Hajun Lv, Kai Lv, Li Ma, Runyuan Ma, Zerun Ma, Wenchang Ning, Linke Ouyang, Jiantao Qiu, Yuan Qu, Fukai Shang, Yunfan Shao, Demin Song, Zifan Song, Zhihao Sui, Peng Sun, Yu Sun, Huanze Tang, Bin Wang, Guoteng Wang, Jiaqi Wang, Jiayu Wang, Rui Wang, Yudong Wang, Ziyi Wang, Xingjian Wei, Qizhen Weng, Fan Wu, Yingtong Xiong, Chao Xu, Ruiliang Xu, Hang Yan, Yirong Yan, Xiaogui Yang, Haochen Ye, Huaiyuan Ying, Jia Yu, Jing Yu, Yuhang Zang, Chuyu Zhang, Li Zhang Pan Zhang, Peng Zhang, Ruijie Zhang, Shuo Zhang, Songyang Zhang, Wenjian Zhang, Wenwei Zhang, Xingcheng Zhang, Xinyue Zhang, Hui Zhao, Qian Zhao, Xiaomeng Zhao, Fengzhe Zhou, Zaida Zhou, Jingming Zhuo, Yicheng Zou, Xipeng Qiu, Yu Qiao and Dahua Lin. 2024. Internl m2 technical report. Preprint, arXiv:2403.17297.  
+Maosong Cao, Taolin Zhang, Mo Li, Chuyu Zhang, Yunxin Liu, Haodong Duan, Songyang Zhang, and Kai Chen. 2025. Condor: Enhance llm alignment with knowledge-driven data synthesis and refinement. arXiv preprint arXiv:2501.12273.  
+Jimmy Carter. 2024. Textocr-gpt4v. https://huggingface.co/datasets/jimmycarter/textocr-gpt4v.  
+George Casella, Christian P Robert, and Martin T Wells. 2004. Generalized accept-reject sampling schemes. Lecture notes-monograph series, pages 342-347.  
+Guiming Hardy Chen, Shunian Chen, Ruifei Zhang, Junying Chen, Xiangbo Wu, Zhiyi Zhang, Zhihong Chen, Jianquan Li, Xiang Wan, and Benyou Wang. 2024a. Allava: Harnessing gpt4v-synthesized data for a lite vision-language model. Preprint, arXiv:2402.11684.  
+Lin Chen, Jisong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. 2023a. Sharegpt4v: Improving large multimodal models with better captions. arXiv preprint arXiv:2311.12793.  
+Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi
+
+Hu, Jiapeng Luo, Zheng Ma, Ji Ma, Jiaqi Wang, Xiaoyi Dong, Hang Yan, Hewei Guo, Conghui He, Botian Shi, Zhenjiang Jin, Chao Xu, Bin Wang, Xingjian Wei, Wei Li, Wenjian Zhang, Bo Zhang, Pinlong Cai, Licheng Wen, Xiangchao Yan, Min Dou, Lewei Lu, Xizhou Zhu, Tong Lu, Dahua Lin, Yu Qiao, Jifeng Dai, and Henhai Wang. 2024b. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. Preprint, arXiv:2404.16821.  
+Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. 2023b. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238.  
+Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. 2023. Vicuna: An opensource chatbot impressing gpt-4 with  $90\%$  * chatgpt quality.  
+Ning Ding, Yulin Chen, Bokai Xu, Yujia Qin, Zhi Zheng, Shengding Hu, Zhiyuan Liu, Maosong Sun, and Bowen Zhou. 2023. Enhancing chat language models by scaling high-quality instructional conversations. arXiv preprint arXiv:2305.14233.  
+Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Xilin Wei, Songyang Zhang, Haodong Duan, Maosong Cao, Wenwei Zhang, Yining Li, Hang Yan, Yang Gao, Xinyue Zhang, Wei Li, Jingwen Li, Kai Chen, Conghui He, Xingcheng Zhang, Yu Qiao, Dahua Lin, and Jiaqi Wang. 2024. Internlm-xcomposer2: Mastering free-form text-image composition and comprehension in vision-language large model. arXiv preprint arXiv:2401.16420.  
+Alexey Dosovitskiy. 2020. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929.  
+Haodong Duan, Junming Yang, Yuxuan Qiao, Xinyu Fang, Lin Chen, Yuan Liu, Xiaoyi Dong, Yuhang Zang, Pan Zhang, Jiaqi Wang, et al. 2024. Vlmealkit: An open-source toolkit for evaluating large multi-modality models. In Proceedings of the 32nd ACM international conference on multimedia, pages 11198-11201.  
+Yann Dubois, Balázs Galambosi, Percy Liang, and Tatsunori B Hashimoto. 2024. Length-controlled alpacaeval: A simple way to debias automatic evaluators. arXiv preprint arXiv:2404.04475.  
+Tinglei Feng, Yingjie Zhai, Jufeng Yang, Jie Liang, Deng-Ping Fan, Jing Zhang, Ling Shao, and Dacheng Tao. 2022. Ic9600: a benchmark dataset for automatic image complexity assessment. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(7):8577-8593.
+
+Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. 2017. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913.  
+Shuhao Gu, Jialing Zhang, Siyuan Zhou, Kevin Yu, Zhaohu Xing, Liangdong Wang, Zhou Cao, Jintao Jia, Zhuoyi Zhang, Yixuan Wang, et al. 2024. Infinity-mm: Scaling multimodal performance with large-scale and high-quality instruction data. arXiv preprint arXiv:2410.18558.  
+Drew A Hudson and Christopher D Manning. 2019. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709.  
+Aniruddha Kembhavi, Mike Salvato, Eric Kolve, Minjoon Seo, Hannaneh Hajishirzi, and Ali Farhadi. 2016. A diagram is worth a dozen images. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 235-251. Springer.  
+Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. 2023. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026.  
+Hugo Laurençon, Andres Marafioti, Victor Sanh, and Léo Tronchon. 2024. Building and better understanding vision-language models: insights and future directions. Preprint, arXiv:2408.12637.  
+Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. 2024a. Llavaonevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326.  
+Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. 2022. Blip: Bootstrapping language-image pretraining for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR.  
+Juzheng Li, Hang Su, Jun Zhu, Siyu Wang, and Bo Zhang. 2018. Textbook question answering under instructor guidance with memory networks. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3655-3663.  
+Tianle Li, Wei-Lin Chiang, Evan Frick, Lisa Dunlap, Tianhao Wu, Banghua Zhu, Joseph E Gonzalez, and Ion Stoica. 2024b. From crowdsourced data to high-quality benchmarks: Arena-hard and benchbuilder pipeline. arXiv preprint arXiv:2406.11939.  
+Xuechen Li, Tianyi Zhang, Yann Dubois, Rohan Taori, Ishaan Gulrajani, Carlos Guestrin, Percy Liang, and
+
+Tatsunori B. Hashimoto. 2023. Alpacaeval: An automatic evaluator of instruction-following models. https://github.com/tatsu-lab/alpaca_eval.  
+Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. 2014. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer.  
+Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. 2023a. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744.  
+Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. 2024a. Llavanext: Improved reasoning,OCR, and world knowledge.  
+Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023b. Visual instruction tuning. arXiv preprint arXiv:2304.08485.  
+Xiao Liu, Xuanyu Lei, Shengyuan Wang, Yue Huang, Zhuoer Feng, Bosi Wen, Jiale Cheng, Pei Ke, Yifan Xu, Weng Lam Tam, et al. 2023c. Alignbench: Benchmarking chinese alignment of large language models. arXiv preprint arXiv:2311.18743.  
+Yang Liu, Yuanshun Yao, Jean-Francois Ton, Xiaoying Zhang, Ruocheng Guo, Hao Cheng, Yegor Klochkov, Muhammad Faaiz Taufiq, and Hang Li. 2024b. Trustworthy llms: a survey and guideline for evaluating large language models' alignment. Preprint, arXiv:2308.05374.  
+Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. 2023d. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281.  
+Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. 2024c. Mmbench: Is your multi-modal model an all-around player? In European conference on computer vision, pages 216-233. Springer.  
+Yuliang Liu, Zhang Li, Hongliang Li, Wenwen Yu, Mingxin Huang, Dezhi Peng, Mingyu Liu, Mingrui Chen, Chunyuan Li, Lianwen Jin, et al. 2023e. On the hidden mystery ofOCR in large multimodal models. arXiv preprint arXiv:2305.07895.  
+Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, KaiWei Chang, Michel Galley, and Jianfeng Gao. 2023. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. arXiv preprint arXiv:2310.02255.  
+Yujie Lu, Dongfu Jiang, Wenhu Chen, William Yang Wang, Yejin Choi, and Bill Yuchen Lin. 2024. Wildvision: Evaluating vision-language models in
+
+the wild with human preferences. arXiv preprint arXiv:2406.11069.  
+Gen Luo, Xue Yang, Wenhan Dou, Zhaokai Wang, Jifeng Dai, Yu Qiao, and Xizhou Zhu. 2024. Monointernvl: Pushing the boundaries of monolithic multimodal large language models with endogenous visual pre-training. arXiv preprint arXiv:2410.08202.  
+Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. 2022. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244.  
+Minesh Mathew, Viraj Bagal, Ruben Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. 2022. Infographicvqa. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1697-1706.  
+Meta. 2024. Llama3.2-vision.  
+OpenAI. 2023a. Chatgpt. https://openai.com/blog/chatgpt.  
+OpenAI. 2023b. Gpt-4 technical report. ArXiv, abs/2303.08774.  
+OpenBMB. 2024. Minicpm: Unveiling the potential of end-side large language models.  
+Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. 2015. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In Proceedings of the IEEE international conference on computer vision, pages 2641-2649.  
+Yusu Qian, Hanrong Ye, Jean-Philippe Fauconnier, Peter Grasch, Yinfei Yang, and Zhe Gan. 2024. Mia-bench: Towards better instruction following evaluation of multimodal llms. arXiv preprint arXiv:2407.01509.  
+Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR.  
+Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. 2024. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36.  
+Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. 2021. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114.
+
+Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. 2018. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of ACL.  
+Wenhao Shi, Zhiqiang Hu, Yi Bin, Junhua Liu, Yang Yang, See-Kiong Ng, Lidong Bing, and Roy Ka-Wei Lee. 2024. Math-llava: Bootstrapping mathematical reasoning for multimodal large language models. arXiv preprint arXiv:2406.17294.  
+Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. 2023. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805.  
+WikiArt Volunteer Team. 2025. Wikiart.  
+Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, et al. 2024. Cambrian-1: A fully open, vision-centric exploration of multimodal llms. arXiv preprint arXiv:2406.16860.  
+Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. 2023. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288.  
+Guan Wang, Sijie Cheng, Xianyuan Zhan, Xiangang Li, Sen Song, and Yang Liu. 2023. Openchat: Advancing open-source language models with mixed-quality data. arXiv preprint arXiv:2309.11235.  
+Weiyun Wang, Zhe Chen, Wenhai Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Jinguo Zhu, Xizhou Zhu, Lewei Lu, Yu Qiao, et al. 2024. Enhancing the reasoning ability of multimodal large language models via mixed preference optimization. arXiv preprint arXiv:2411.10442.  
+Zhangchen Xu, Fengqing Jiang, Luyao Niu, Yuntian Deng, Radha Poovendran, Yejin Choi, and Bill Yuchen Lin. 2024. Magpie: Alignment data synthesis from scratch by prompting aligned llms with nothing. arXiv preprint arXiv:2406.08464.  
+Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. 2023. Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490.  
+Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. 2023. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. arXiv preprint arXiv:2311.16502.
+
+Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, et al. 2024. Recognize anything: A strong image tagging model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1724-1732.  
+Xiangyu Zhao, Yicheng Chen, Shilin Xu, Xiangtai Li, Xinjiang Wang, Yining Li, and Haian Huang. 2024. An open and comprehensive pipeline for unified object grounding and detection. arXiv preprint arXiv:2401.02361.
+
+# A Alignment of MLLMs with Human Preference
+
+Although open-source MLLMs have already matched or even surpassed proprietary models like the GPT and Claude series in common VQA tasks like OCR and Visual Perception, a significant gap remains in their alignment with human preferences. When posed with open-ended questions that require knowledge-rich responses, even the most advanced open-source MLLM, InternVL2-76B, struggles to provide comprehensive answers with high readability, as illustrated in Fig. 6. In contrast, GPT-4o not only accurately identifies the main objects relevant to the question but also provides well-structured responses enriched with comprehensive contextual knowledge, achieving a high level of alignment with human preferences.
+
+# B Image sources
+
+We carefully select image sources for Arts, Charts, Diagrams, and Posters tasks. For Arts, WikiArt (Team, 2025) is selected as the image source, offering a diverse range of painting styles. We uniformly sample 2000 images across all painting styles to ensure diversity in the dataset. For Charts, we select ChartQA (Masry et al., 2022), a dataset featuring charts that contain substantial statistical information. ChartQA includes several subcategories, from which we filter out simplistic charts with only two columns and retain those that contain charts with rich contextual information and diverse types. For Diagrams, we choose images from TextbookVQA (Team, 2025), which provides diagrams rich in natural content and detailed explanations. We exclusively utilize the question images and teaching images from the image sources, as they meet our specific requirements. For Posters, we utilize InfographicsVQA (Mathew et al., 2022), a dataset containing high-quality posters with intricate designs and informative content.
+
+# C Chart Post-Refine Pipeline
+
+After curating images along with high-quality question, we first employ the current most powerful models(GPT-4o, InternVL2-72B, Qwen2VL-72B) to generate answers separately. A powerful LLM(Qwen2.5-72B) is then utilized to extract objective facts from the chart within each generated answer. The facts extracted from the answers of multiple models are compared for consistency. If
+
+(a) Overly Simplistic Question & Answer  
+![](images/25eda9114e522878608754919d52e0f09124b5164a87c64aefb6e7df0d32dc75.jpg)  
+Q: What animal is the main subject of the image?  
+A: The main subject of the image is a young sheep or lamb.
+
+![](images/9a9c2b3d300cf04c4a93e27869bc8863691ca199c3c046264a1a4b84bdc70e3d.jpg)  
+Q: Based on the design elements observed, what genre of music is this guitar likely designed to play?  
+A. Considering the stylized and modern look of the guitar with sharp angles and cutaways, it is likely designed for genres that often utilize electric guitars with a more edgy aesthetic, such as rock, metal, or punk music. These genres typically favor electric guitars with distinctive shapes and designs that stand out visually during performances.
+
+(b) Semantically Impoverished Image  
+(c) Ambiguous Question  
+Figure 5: Examples of limitation with current multimodal instruction tuning dataset.  
+![](images/82c32bb4786b30142066c0a7d500a207c4f420ff65afa1e28b130efb2c8f45e0.jpg)  
+Q: What might be the purpose of the toy cow in relation to the cup with the contest promotion?  
+A: Based on the image, the toy cow may be a playful addition to the scene, possibly referencing the dairy aspect of the product that might be associated with the cup (for example, if the cup was from a coffee chain, the cow could be a nod to the use of milk in coffee). It's not clear if there's an intentional relationship, but it could also be part of a personal collection or a decorative item in the setting.
+
+the facts differ significantly and lead to entirely different conclusions, such responses are flagged for further review or discarded to avoid misinformation. For cases where the facts exhibit only minor differences, we merge the detailed factual content from Qwen2VL-72B into the comprehensive explanations provided by GPT-4o. The merged answers are further reviewed and supervised by human experts to ensure their quality and consistency.
+
+# D Training Details
+
+Our training strategy largely follows the approaches adopted by LLaVA and LLaVANext. CLIP-Large-336-Patch14 is employed as the visual encoder. In line with the LLaVA training strategy, we first conduct a pretraining stage where both the visual encoder and the LLM are frozen. We utilize the LLaVA-pretrain-558k and ALLaVA-pretrain-728k datasets for pretraining. The batch size is uniformly set to 256 and learning rate is set to 1e-3. During this phase, images are resized to  $336 \times 336$ , and no image-splitting method is applied.
+
+For SFT stage, we unfreeze the LLM for LLaVA and further unfreeze the visual encoder for LLaVANext. In the case of LLaVANext, we apply the image-splitting method, setting the maximum split size to  $3 \times 3$ . The batch size is uniformly set to 128 and learning rate is set to 2e-5.. LLaVAInternLM2.5-7B is trained using  $8 \times \mathrm{A}800$  GPUs for 12 hours. LLaVANext-InternLM2.5-7B is trained using  $16 \times \mathrm{A}800$  GPUs for 13 hours. LLaVANext-Qwen2.5-32B is trained using  $32 \times \mathrm{A}800$  GPUs for 24 hours.
+
+<table><tr><td>Model</td><td>MM-AlignBench</td><td>WildVision</td><td>MMVet</td></tr><tr><td>LLaVANext-77k</td><td>15.1/-52.6</td><td>13.6 / -63.1</td><td>37.7</td></tr><tr><td>+ 33k w.o. Imager Filter</td><td>31.4 / -42.3</td><td>22.0 / -42.3</td><td>42.0</td></tr><tr><td>+ 33k w. Imager Filter</td><td>35.3 / -41.0</td><td>22.6 / -37.5</td><td>44.4</td></tr></table>
+
+Table 8: Ablation study on the impact of utilizing image filter.
+
+# E Ablation on Image Filter
+
+We conduct ablation studies to evaluate the impact of using our image filter. We randomly sample 77k images from the LLaVA-Next-778K SFT dataset as a baseline and separately select images from CC3M both with and without applying our image filter. Subsequently, the selected images are used to generate Knowledge and Inferential Question-Answer pairs following the pipeline described in Sec. 4.3. We then assess the effect of these datasets on the performance of LLaVA-Next-InternLM2.5-7B, with the results presented in Tab. 8. It can be observed that by utilizing the image filter, the selected images contain richer semantic information, leading to the generation of higher-quality image-question pairs. This, in turn, enhances the model's performance in terms of alignment with human preferences.
+
+# F Human Experts
+
+In this study, two authors (both PhD students in computer science) serve as human experts. They are responsible for reviewing and refining the questions in MM-AlignBench, as well as evaluating and filtering incorrect merged cases following the Chart Post-Processing Refinement stage.
+
+# G License
+
+The InternLM and Qwen models are licensed under the Apache-2.0 license. The ChartQA dataset is distributed under the GNU General Public License v3.0. The remaining datasets are licensed under CC BY-NC 4.0, which permits only non-commercial use.
+
+![](images/8fbcab58c750be139cc1ed096811be73418f79463d1c1892e4e6bb8ff3235794.jpg)  
+Figure 6: GPT-4o shows superior alignment with human preference than InternVL2-76B.
+
+# Text Prompt
+
+How does the loading and unloading process depicted in the image contribute to international shipping and trade?
+
+What are the potential logistical challenges faced at such ports?
+
+# InternVL2-76B:
+
+Ambiguous
+
+Ignorance
+
+Simplistic
+
+The loading and unloading process shown in the image is a critical component of international shipping and trade, as it facilitates the movement of goods between different countries and regions. The use of cranes and trucks allows for the efficient transfer of cargo from ships to land transport, and vice versa. This process is essential for the global supply chain, enabling the distribution of goods to various markets. Potential logistical challenges at such ports include coordinating the arrival and departure of ships, managing the flow of trucks and other vehicles, ensuring the safety and security of cargo, and dealing with weather conditions that can impact operations. Additionally, there may be challenges related to customs clearance, documentation, and compliance with international trade regulations. The image suggests a well-organized operation, but these challenges are inherent to the complex and dynamic environment of international shipping and trade.
+
+# GPT-4o:
+
+Clarity
+
+Knowledge-Rich
+
+Comprehensive
+
+The loading and unloading process depicted in the image is a critical component of international shipping and trade. This process involves transferring goods between ships and other forms of transportation, such as trucks, which facilitates the movement of goods across different regions and countries. Here's how it contributes to international shipping and trade, along with potential logistical challenges:
+
+# Contribution to International Shipping and Trade:
+
+# 1.Efficient Movement of Goods:
+
+The process ensures that goods are efficiently transferred from ships to trucks or vice versa, enabling the seamless flow of products from manufacturers to consumers across the globe.
+
+# 2.Intermodal Transportation:
+
+Ports serve as hubs for intermodal transportation, where goods are transferred between different modes of transport (e.g., sea to land). This interconnectivity is essential for global supply chains.
+
+# 3.Economic Growth:
+
+Ports facilitate international trade, which is a significant driver of economic growth. Efficient port operations can enhance a country's trade capabilities and competitiveness.
+
+# 4. Supply Chain Management:
+
+Effective loading and unloading processes are crucial for maintaining the integrity and timing of supply chains, ensuring that goods reach their destinations on schedule.
+
+# Potential Logistical Challenges:
+
+# 1.Congestion:
+
+Ports can become congested, especially during peak times, leading to delays in loading and unloading. This can disrupt supply chains and increase costs.
+
+# 2.Infrastructure Limitations:
+
+Inadequate infrastructure, such as insufficient cranes, storage facilities, or transportation links, can hinder the efficiency of port operations.
+
+# 3.Customs and Regulatory Compliance:
+
+Navigating customs and regulatory requirements can be complex and time-consuming, potentially causing delays in the movement of goods.
+
+# 4.Labor Issues:
+
+Labor shortages, strikes, or disputes can impact port operations, leading to delays and increased costs.
+
+# 5.Weather Conditions:
+
+Adverse weather conditions can disrupt port activities, affecting the loading and unloading process and causing delays.
+
+# 6.Security Concerns:
+
+Ensuring the security of goods and preventing theft or damage is a significant challenge. Ports must implement robust security measures to protect cargo.
+
+# 7.Technological Integration:
+
+Integrating advanced technologies for tracking, managing, and automating port operations can be challenging but is necessary for improving efficiency and reducing errors.
+
+# 8.Environmental Regulations:
+
+Ports must comply with environmental regulations, which can impact operations and require investments in cleaner technologies and practices. Addressing these challenges requires coordinated efforts between port authorities, shipping companies, logistics providers, and government agencies to ensure smooth and efficient port operations, which are vital for sustaining international trade and economic growth.
+
+![](images/4ea118d7d3a28cf57233f3a3685f1d06c04f62d6cf791ce97d26f51e232ed5ae.jpg)  
+IC Score=0.10  
+Low IC
+
+![](images/d8fde4bb719b3076f09ed64441402e74385e504849d7a7033eb493f95c976caf.jpg)  
+IC Score=0.89  
+High IC  
+RAMCount=6  
+Meaningless
+
+![](images/a411aa8fb1b3dcaf1074b8972a67a3f7a84c3a6c4991b5e57370d8e2a40134f4.jpg)  
+IC Score=0.69  
+High IC  
+RAMCount=12
+
+# High Content Redundancy
+
+![](images/75e8a2281e2ad07f07924e81767ae86545637350dd2b6b32ab50b83edfe4d467.jpg)  
+Low RAM Count  
+Figure 7: Demonstration examples of our image filter.
+
+# Rich Semantic Information
+
+![](images/8becb52416f4ec6be9b5d6f59586ff6ec72d876b8813a01a2cb94b7a9114a836.jpg)  
+High RAM Count
+
+![](images/7c87ab89c82809726c4619017df0da2f749643025012a6a82804c21b050771b2.jpg)
+
+# Text Prompt:
+
+How many chairs are in this pic?
+
+Reference Answer:
+
+The image shows 4 chairs in what appears to be a waiting or reception area outside an "H&R Block" office or service location.
+
+# Simplistic & Factual Question
+
+![](images/89eb2c5a86ab729b891f6488e6379312f8892c4c8833d8da15b856d8d58ba4e4.jpg)
+
+# Text Prompt:
+
+What is this?
+
+Reference Answer:
+
+The image depicts a dense forest of coniferous trees, likely pines or firs, characterized by their tall, slender trunks and conical shapes. The trees are uniformly green, suggesting they are healthy and possibly located in a temperate climate where such trees thrive. The forest appears lush and undisturbed, indicating it may be a protected natural area or a well-managed forest. The perspective is from ground level, looking upwards into the canopy, which emphasizes the height and density of the trees. There are no visible signs of human activity or structures, reinforcing the impression of a pristine, natural environment.
+
+# Meaningless image & Question
+
+![](images/b206531a04eea486bbe2fe3ea826873c98675cb68878eca8f821d44d5dbe6444.jpg)  
+Figure 8: Examples of limitation within current multi-modal benchmark for alignment.
+
+# Text Prompt:
+
+a chef standing in a kitchen, pizza's recepie such as flower and tomatoes are on top of the shelves, a pizza on the background being cooked inside an oven
+
+Reference Answer:
+
+The image does not depict a chef standing in a kitchen with pizza ingredients or an oven cooking pizza as you described. Instead, the image shows a close-up portrait of an older man with gray hair and wrinkled features. He is wearing a white chef's uniform or jacket and has a serious, slightly stern expression on his face. The background is a solid dark gray or black color, putting the focus entirely on the man's face and upper body.
+
+# Confusing Prompt
+
+# Knowledge
+
+![](images/9b30dca2ba22afd12165f42cc2aa9c009472b93bcaddcaaa1e822c00aad0c29c.jpg)
+
+# Please introduce the history of the landmark in the image, and suggest a culturally enriching leisure activity plan for visitors.
+
+History of the Landmark
+
+The landmark in the image is the Colosseum, an iconic symbol of ancient Rome located in the heart of the city. Constructed between 70-80 AD under the reign of Emperor Vespasian and completed by his successor Titus, the Colosseum was the largest amphitheater ever built at the time and could hold between 50,000 to 80,000 spectators.
+
+It was primarily used for gladiatorial contests and public spectacles such as animal hunts, mock sea battles, and executions. Despite suffering damage from natural disasters and stone robbers, the Colosseum remains a testament to Roman engineering and architectural prowess.
+
+Culturally Enriching Leisure Activity Plan
+
+Moming:Guided Historical Tour
+
+Start your day with a guided tour of the Colosseum. A knowledgeable guide will provide insights into the architectural techniques used in its construction and recount tales of the gladiators who fought within its walls. This tour will help visitors appreciate the historical significance and the cultural heritage of the Colosseum.
+
+### Afternoon: Roman Forum Exploration
+
+After the Colosseum, head to the nearby Roman Forum, once the center of Roman public life. Walk through the ruins and imagine the political, religious, and economic activities that took place here. Consider using an audio guide or a historian-led tour for deeper understanding.
+
+Evening: Traditional Italian Dining Experience
+
+Conclude your day with a dinner at a nearby trattoria, indulging in traditional Italian cuisine. Opt for dishes such as pasta alla carbonara or cacio e pepe, paired with a glass of local wine. End your meal with a classic tiramisu or gelato, savoring the flavors of Rome.
+
+This plan not only immerses visitors in the rich history of the Colosseum and its surroundings but also offers a taste of authentic Roman culture.
+
+# Inferential
+
+![](images/0597876a6964e00c10e2bd15c1247bb7d380c4bffd62ec5e909c7bd16184a023.jpg)
+
+# What occasion are the people in the image likely celebrating?
+
+The people in the image are likely celebrating a holiday, most probably Christmas. This conclusion is drawn from several key elements visible in the scene:
+
+- Wrapped Gifts: The presence of wrapped gifts suggests a gift-giving occasion, which is a common tradition during Christmas.  
+- Stockings on the fireplace: Stockings hung by the fireplace are a traditional Christmas decoration, often used for Santa Claus to fill with small gifts and treats.  
+- Festive Decorations: The overall festive atmosphere, including other possible decorations like a Christmas tree or lights, further supports the idea that this is a Christmas celebration.
+
+These elements together create a clear picture of a family or group gathered to enjoy the warmth and joy of the holiday season.
+
+# Creation
+
+![](images/97945c5c0f5c5a34c9d81dbf99dd56bf7efdcc7176278425706e294b44326c35.jpg)  
+Figure 9: Examples of each task in OmniAlign-V.
+
+# Imagine you are the athlete in this image. Describe your thoughts and emotions as you prepare for an important game. What motivates you to give your best performance on the ice?
+
+Standing by the rink, I feel the electric buzz of anticipation coursing through my veins. The cool air of the arena sharpens my focus, while the roar of the crowd fuels the fire within me. Wearing the Vancouver jersey, I am more than just another athlete on the ice; I am part of something greater, a legacy I am determined to uphold.
+
+As I grip my stick, memories of early mornings on frozen ponds flash through my mind. Each moment, each sacrifice, has led me to this pivotal game. My heart beats in sync with the rhythm of skates slicing through ice, a reminder that this is where I belong. The motivation to perform doesn't just come from a desire to win, but from the faces of my teammates who have become family, the unwavering support from fans, and the pride in representing my city. I visualize each play, every pass, every shot, letting the anticipation transform into confidence.
+
+Tonight, as the whistle blows and the puck drops, I am ready to give everything I have. This is my stage, my moment, and I intend to seize it with every ounce of passion in me.
+
+# Chart
+
+Coal reserves, 2019
+
+Proved reserves is generally taken to be those quantities that geological and engineering information indicates with reasonable certainty can be recovered in the future from known reservoirs under existing economic and operating conditions.
+
+![](images/1fcec9cf227df9f493d945a51f371330e40302c463cac162dd510ff63e448ebf.jpg)
+
+![](images/8e32e058ef9ebb348016d265e46d4114f36198fb27b34007c3783a3f5593459c.jpg)  
+Figure 10: Examples of each task in OmniAlign-V.
+
+# What is the relationship between the coal reserves of Brazil, Czechia, and Uzbekistan, and how might this affect their energy policies?
+
+The coal reserves of Brazil, Czechia, and Uzbekistan differ significantly, which may influence their energy policies and strategies:
+
+1. Coal Reserves Overview
+
+- Brazil: 6.6 billion tons (largest reserves among the three).
+
+- Czechia: 2.93 billion tons (moderate reserves).
+
+- Uzbekistan: 1.38 billion tons (smallest reserves).
+
+##2. Implications for Energy Policies
+
+- Brazil:
+
+- Its large coal reserves might provide a reliable domestic energy source for power generation and industry.
+
+- However, Brazil has a strong focus on renewable energy (e.g., hydropower) and may prefer to limit coal use to reduce environmental impacts.
+
+- Coal could be used as a backup or exported for economic gains.  
+- Czechia:
+
+- With moderate reserves, coal has traditionally played a key role in Czechia's energy mix, particularly for electricity generation.
+
+- Czechia faces pressure to transition away from coal due to European Union climate targets, potentially accelerating the shift toward renewables and nuclear energy.
+
+-Uzbekistan:
+
+- Limited coal reserves may constrain its reliance on coal for energy.
+
+- Uzbekistan has been investing in natural gas and renewable energy, and its smaller coal reserves may push it to focus more on diversifying its energy sources.
+
+3. Global Transition Trends
+
+- All three countries may face international pressure to reduce coal use in favor of cleaner energy sources due to global climate change goals.
+
+- The extent of their reliance on coal will depend on domestic policies, economic priorities, and access to alternative energy resources.
+
+4. Export Potential
+
+- Brazil, with the largest reserves, has more potential for coal exports, which could shape its energy trade policies.
+
+- Czechia and Uzbekistan are less likely to focus on coal exports due to smaller reserves and domestic needs.
+
+In summary, Brazil's larger reserves provide more flexibility for domestic use and export, while Czechia and Uzbekistan's smaller reserves may push them to prioritize energy diversification and transition to cleaner energy.
+
+# Diagram
+
+![](images/5589c3b09ed4e112e0ac9f44e46892bb7b04317b2136d06c1be37e1bc9b48123.jpg)
+
+# How does the relationship between weathering and sediment formation affect the rock cycle?
+
+The relationship between weathering and sediment formation is a vital component of the rock cycle, as it provides the materials necessary for the formation of sedimentary rocks. Here's how they interact and affect the rock cycle:
+
+1. Weathering Breaks Down Rocks:
+
+- Weathering is the process in which rocks are broken down into smaller particles, such as sediments, through physical, chemical, or biological means.
+
+- Rocks exposed at the Earth's surface, including igneous, metamorphic, or even sedimentary rocks, are subjected to weathering.
+
+2. Sediment Formation and Transportation:
+
+- The weathered material becomes sediments, which are transported by wind, water, or ice to new locations. This transportation step is essential for distributing sediments across different environments.
+
+3. Deposition and Lithification:
+
+- Sediments are eventually deposited in layers, often in water bodies like rivers, lakes, or oceans. Over time, these sediments are compacted and cemented through lithification to form sedimentary rocks.
+
+4. Sedimentaryary Rocks and the Rock Cycle:
+
+- Sedimentary rocks can undergo further changes in the rock cycle:
+
+- They may be subjected to heat and pressure to become metamorphic rocks.
+
+- They can also be subducted into Earth's mantle, where they melt and form magma, which solidifies into igneous rocks.
+
+5. Continuous Cycling:
+
+- Weathering and sediment formation ensure that materials from older rocks re-enter the rock cycle, highlighting the interconnectedness of Earth's processes.
+
+In summary, weathering and sediment formation are essential steps in recycling materials within the rock cycle, connecting the processes that form sedimentary, metamorphic, and igneous rocks.
+
+![](images/d84466bd3c7482e36d50bea68f9e959366422a3fed635bd43aa85af8dd3c9a99.jpg)  
+Figure 11: Examples of each task in OmniAlign-V.
+
+# In what ways does the infographic suggest small businesses can differentiate themselves from larger competitors, and what strategies are highlighted as crucial?
+
+The infographic highlights several ways small businesses can differentiate themselves from larger competitors and underscores specific strategies as crucial for growth:
+
+Ways to Differentiate:
+
+1. Agility in Business Strategy:
+
+-  $21\%$  of small businesses are changing their business strategies to effectively compete with larger companies.
+
+- The focus is on flexibility and responsiveness to market changes.
+
+2. Expanding Product and Service Offerings:
+
+- 1 in 5 small businesses now offer additional products or services to attract and retain clients.
+
+3. Geographical Expansion:
+
+-  $37\%$  of SMEs plan to expand into new areas over the coming year to tap into untapped markets.
+
+4. Customer Relationships:
+
+-  $69\%$  prioritize winning new customers and partners.
+
+-  $58\%$  focus on strengthening relationships with existing customers and partners.
+
+Crucial Strategies:
+
+1. Leveraging Technology:
+
+- 65% of SMEs recognize that technology improves communication with customers.
+
+-  $24\%$  are prioritizing investment in technology (e.g., mobile hardware and security systems).
+
+2. Social Media and Networking:
+
+-  $58\%$  see social tools as essential for operations.
+
+- 2 in 5 SMEs plan to use mobile social networking to connect with other small businesses.
+
+3. Community and Support:
+
+-  $63\%$  find advice and support from other small businesses helpful, emphasizing the importance of collaboration and shared insights.
+
+Key Takeaway:
+
+The infographic emphasizes that investing in technology, adapting strategies, and building strong customer relationships are critical for small businesses to differentiate themselves and thrive in competitive markets.
+
+![](images/3197bc4871d8b287f88877cef482180951beba6fa2e3ceba5a2734ad7f6ce8ce.jpg)  
+Art
+
+What are the prominent features of the setting and figures depicted in the painting, and how do they reflect the themes and techniques commonly used by the artist?
+
+The painting 'The Charmes of Life' by Antoine Watteau showcases a vivid and lively scene set in a pastoral landscape, typical of Watteau's work. Watteau is known for his 'fêtes galantes' style, which depicts scenes of aristocratic outdoor entertainment. ### Features and Themes:
+
+- Setting: The painting features a lush, picturesque countryside, suggesting a leisurely and idyllic lifestyle. The use of soft brushwork and a subtle color palette creates an ethereal, dreamlike quality.  
+- Figures: The elegantly dressed figures are engaged in music and conversation, portraying a sense of enjoyment and harmony. The attire and gestures reflect the social elegance of the Rococo period.  
+- Architectural Elements: The grand columns and structured terrace highlight the blend of nature and refined culture, a common theme in Watteau's works.
+
+# Techniques:
+
+- Color and Light: Watteau's use of light captures the luminous quality of the outdoors, enhancing the serene and carefree atmosphere.
+
+- Composition: The arrangement of figures in a relaxed, informal manner typifies the fluidity and spontaneity characteristic of Rococo art.
+
+Overall, the painting reflects Watteau's mastery in capturing the fleeting pleasures of life, combining sophisticated elegance with a deep appreciation for nature and human interaction.
+
+![](images/f14f734fe9b878b8647fdfa54fbe143068f6f2d5ce56123b3aeb0b8e3403cbf1.jpg)  
+Instruction Following  
+Figure 12: Examples of each task in OmniAlign-V.
+
+What elements in the image reflect urban fashion trends, and how do these elements align with current fashion industry predictions for seasonal attire? Please describe each element using a metaphor and ensure your response includes at least three distinct metaphors.
+
+The image showcases several elements reflective of urban fashion trends, each of which can be described using vivid metaphors to highlight their significance and alignment with current fashion industry predictions for seasonal attire.
+
+1. Floral Dress: The central figure's floral dress is like a garden in full bloom, with its bright, bold patterns that capture the essence of spring and summer. This aligns with fashion industry predictions that emphasize vibrant colors and nature-inspired designs as key trends for these seasons. Just as a garden bursts with life and color, the dress brings a sense of freshness and vitality to the urban landscape.  
+2. Bright Pink Shoes: The bright pink shoes are like a neon sign in a bustling city, drawing attention and adding a pop of color to the outfit. This technique of using bold, eye-catching colors is a common strategy in urban fashion to make outfits more dynamic and memorable. The shoes serve as a beacon of style, guiding the eye and adding a modern, urban flair to the ensemble.  
+3. Bright Orange Handbag: The bright orange handbag carried by the person on the left is like a vibrant sunset, blending practicality with style. It complements the idea of combining functional fashion items with bold, statement pieces, a growing trend in urban fashion. Just as a sunset transforms the sky with its warm, radiant hues, the handbag adds a touch of warmth and energy to the overall look, making it both practical and visually striking.
+
+These elements collectively reflect the broader industry move towards integrating bold, statement pieces with functional fashion items that cater to an urban lifestyle, ensuring that the wearer stands out while remaining practical and stylish.
+
+# Prompt for Knowledge Task
+
+Examine the image provided and generate a knowledge-based and exploratory question based on the content of the image and supply corresponding detailed answers.
+
+# Question Guidelines:
+
+- Your question should invite insightful discussion on the types of elements in the image, such as:  
+- **Objects**: For example, animals, plants, food, or products.  
+- E.g., "What breed of dog is in the picture, and what are their characteristics?", "Can you give me a recipe for the food in image?", "Write a Product Description for the product in the image".  
+- **Locations and Features**: Relevant to countries, landmarks, famous people, or scenic spots.  
+- E.g., "Please introduce the history of the landmark in the picture.", "How did the states in image get their names?", "Who is the person in the image and what is him famous for?".  
+- **Activities and Technologies**: Related to sports, machines, technology, and environmental details.  
+- E.g., "Can you explain how the game in image is played?", "How is the machine shown in the image operated?"  
+- **Events, Literature, and Media**: Concerning books, movies, or series in the image.  
+- E.g., "Write a short description about the movie or series in the image.", "Think of books that would be enjoyable for someone who liked the books in the image."
+
+# Answer Guidelines:
+
+- Ensure your answers are factual and comprehensive.  
+- Please use Markdown formatting in your text to enhance the content, making it visually appealing and easy to read. Include appropriate headings, subheadings, lists, code blocks, and other Markdown elements to optimize your answers.
+
+# Output Format:
+
+Your response should strictly follow this format:
+
+```txt
+```
+```
+```
+```
+```
+```
+```
+``
+```
+
+Figure 13: An Example of the prompt for Knowledge task generation.
+
+# Prompt for Creative tasks
+
+You are a skilled writer with a talent for crafting insightful and engaging questions based on the content of a given image.
+
+**Task Guidelines**:
+
+- Analyze the content of the provided image and select one of the question types listed below. Use it to create engaging questions that lead the viewer to explore and interpret the image in different level of complexity.  
+- Be closely tied to the content of the image, emphasizing its primary visual or thematic elements.  
+- Your questions should **avoid directly referencing specific details** in the image. Instead, they should encourage deeper reflection, ensuring the question cannot be answered without seeing the image.  
+- Avoid overly rigid or direct phrasing, focusing instead on open-ended exploration.
+
+**Question Types**
+
+Your questions can be from the following types, each followed by an example with a different level of complexity.
+
+- Simple (basic observation or initial reflection)  
+- Moderate (more thought-provoking, requiring a deeper understanding and more structured response)  
+- Difficult (complex or abstract, requiring analysis and strict formatting)
+
+{Match Types}
+
+**Output Format**:
+
+Your response should strictly follow this format:
+
+```snap
+```
+```
+```
+```
+```
+```
+```
+```
+```
+```
+```
+```
+``
+```
+
+Figure 14: An Example of the prompt for Creative task generation.
+
+# Prompt for Inferential tasks
+
+You are an image analysis expert skilled in posing high-quality **inferential questions**. Please provide 2-5 of the most insightful questions you can think of, following these guidelines:
+
+For Questions:
+
+- **Focus on image-based questions:** Ensure that your question cannot be answered without analyzing the image. **You should not directly provide image's data or details in your generated questions.** For example, "What might be the impact on the radio industry due to 34 stations having their licenses revoked?" includes the data in the image and can be answered without analyzing the image, so it is a bad question.  
+- Ensure that questions are natural, not overly rigid. **You should be quite certain and confident about the questions you pose and their answers, and avoid using words like "possibly", "maybe" or "might be" in both questions and answers.**  
+- Your questions must require reasoning beyond the direct content of the image, making reasonable inferences based on the information presented. - The scope of the questions should not be overly broad or delve into political, philosophical, speculative, sensitive, or controversial topics. Stay within the context of the scene and elements inferred from it.
+
+For Answers:
+
+- You should provide a clear and concise answer to the question.  
+Good Examples:  
+- What precautions are the people on the boat taking to stay comfortable during the trip? - Is there anything else on the table other than the pizza? - Why do these people choose to dress in this style?  
+- What decorative element is present in this public restroom that is not typical? Bad Examples:  
+- What might indicate
+
+Figure 15: An Example of the prompt for Inferential task generation.
+
+# Prompt for Chart tasks
+
+You're a great image analyst. You need to analyze the image provided and generate some insightful questions based on the content of the image.
+
+Question generation guidelines:
+
+- Ensure that your questions require the image to be answered and do not include explicit information from the image. Instead, pose questions that prompt the respondent to analyze the image to find the answer.  
+- Your question should be explainable and require some reasoning to answer.  
+- Your question could contain different analytical perspectives, such as trends, comparisons, causal inference, etc.  
+- - Your questions should be insightful but also clear and straightforward. Avoid overly complex or niche questions.
+
+Bad question examples:
+
+- "What might be some factors contributing to the significantly higher private health expenditure per person in Argentina compared to Fiji and Benin?" (This includes specific details from the chart.) Its correct clarification should be "Is there any difference in private health expenditure per person between Argentina, Fiji, and Benin? If so, what might cause the difference?"  
+- "What trends can be observed in private health expenditures per person among the three countries shown?" (This question is unclear because 'trends between countries' is not a standard analytical concept. Trends typically refer to patterns over time or categories, not direct cross-entity comparisons.)
+
+Output format:
+
+Your response should strictly follow this format:
+
+```txt
+{   
+"questions":[   
+{   
+"question": "Question text"   
+},   
+{   
+"question": "Another question text"   
+}   
+]
+```
+
+Figure 16: An Example of the prompt for Chart task generation.
+
+# Prompt for Poster tasks
+
+You're an excellent image analyst and good at generating insightful questions about the image. You need to analyze the image provided and generate some insightful questions based on the content of the image, and you should answer the questions you generated.
+
+Possible Categories Reference:
+
+1. **Cultural and Social Context**  
+2. **Analysis and Inference**.  
+3. **Visual Elements and Design Techniques**
+
+Question Guidelines:
+
+- **Focus on image-based questions:** Ensure that your question cannot be answered without analyzing the image. **You should not directly provide image's data or details in your generated questions.** For example, "What might be the impact on the radio industry due to 34 stations having their licenses revoked?" includes the data in the image and can be answered without analyzing the image, so it is a bad question.  
+- **Encourage thoughtful, structured responses:** Your question should be explainable and need some reasoning to answer. You should not generate questions that are just extracting information from the image. For example, "What percentage of organizations verify the past employment records of new employees according to the image?" is a bad question.  
+- **Ensure diversity in the questions:** Cover different aspects of the image, encouraging multiple perspectives. You can choose some appropriate categories from the possible categories reference. **For the same category, you can generate multiple questions.**  
+- **Generate high-quality questions:** You can choose to generate challenging questions, but their answers should be able to clearly explain.
+
+Output Format:
+
+Your response should strictly follow this format:
+
+```txt
+"questions": [  
+{  
+"question": "Question text"  
+}  
+]
+```
+
+Figure 17: An Example of the prompt for Diagram task generation.
+
+# Prompt for Diagram tasks
+
+You're a great diagram analyst. You need to analyze the diagram provided and generate 2-4 insightful questions based on the content of the diagram.
+
+- Your questions should be guiding and should not directly point to the content. For example: "How does acid rain affect water bodies, soil, and plant life?" should be changed to "How does the process in image affect water bodies, soil, and plant life?"
+
+Question Guidelines:  
+- Your question should invite insightful discussion, such as:  
+- \*\*Interpretation\*\*: Symbol Interpretation, Data Extraction  
+- **Examples**:  
+- "What is the role of the cytokine-producing cell in the process shown?"  
+- "Enumerate the steps outlined in the flowchart."  
+- \*\*Relation Analysis\*\*:  
+- **Examples**:  
+- "How does variable A affect variable B in the diagram?"  
+- "How many ways can A to B be achieved in the diagram?"  
+- \*\*Inference\*\*:  
+- **Examples**:  
+- "What can be inferred about the system's stability from the diagram?"  
+- "What does the bacterium do once it has the hybrid plasmid?"
+
+Output Format:
+
+Your response should strictly follow this format:
+
+```txt
+```
+```
+```
+```
+```
+```
+``
+```
+
+Figure 18: An Example of the prompt for Diagram task generation.

ACL/2025/OmniAlign-V_ Towards Enhanced Alignment of MLLMs with Human Preference/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5577a98348b51cb82d7a7ec4cc1701ad8871cedcd97095e00dedcc486d20c215
+size 1221544

ACL/2025/OmniAlign-V_ Towards Enhanced Alignment of MLLMs with Human Preference/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7e5296438e2e809f64a3359b6e387281cc259ba4c56ac13e62f6931537d0a49d
+size 847869

ACL/2025/OmniCharacter_ Towards Immersive Role-Playing Agents with Seamless Speech-Language Personality Interaction/6a448e7d-5e2f-40a1-b157-bca2f1ee96c7_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2092080a78d734895b4d9412b4a0832574f6bb241875ba7d47ba3e97e7530275
+size 104323

ACL/2025/OmniCharacter_ Towards Immersive Role-Playing Agents with Seamless Speech-Language Personality Interaction/6a448e7d-5e2f-40a1-b157-bca2f1ee96c7_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:023ead32cfc5453756e9dd33f9b75aba6e3bbeadeff5e13538ab7018dd8d2d92
+size 124701

ACL/2025/OmniCharacter_ Towards Immersive Role-Playing Agents with Seamless Speech-Language Personality Interaction/6a448e7d-5e2f-40a1-b157-bca2f1ee96c7_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e6adf8db64654e94b37ec08836b2c4d9666dda52d50c65fdc3851c8d5608d853
+size 3743187

ACL/2025/OmniCharacter_ Towards Immersive Role-Playing Agents with Seamless Speech-Language Personality Interaction/full.md

@@ -0,0 +1,410 @@
+# OmniCharacter: Towards Immersive Role-Playing Agents with Seamless Speech-Language Personality Interaction
+
+Haonan Zhang $^{1,2*†}$ , Run Luo $^{2,4,5*}$ , Xiong Liu $^{2*}$ , Yuchuan Wu $^{2}$ , Ting-En Lin $^{2}$ , Pengpeng Zeng $^{1}$ , Qiang Qu $^{4,5}$ , Feiteng Fang $^{2,5}$ , Min Yang $^{4,5}$ , Lianli Gao $^{3}$ , Jingkuan Song $^{1‡}$ , Fei Huang $^{2}$ , Yongbin Li $^{2‡}$
+
+$^{1}$ Tongji University  $^{2}$ Tongyi Laboratory  $^{3}$ Independent Researcher  $^{4}$ University of Chinese Academy of Sciences
+
+Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences {zchiowal, jingkuan song}@gmail.com shuide.lyb@alibaba-inc.com
+
+# Abstract
+
+Role-Playing Agents (RPAs), benefiting from large language models, is an emerging interactive AI system that simulates roles or characters with diverse personalities. However, existing methods primarily focus on mimicking dialogues among roles in textual form, neglecting the role's voice traits (e.g., voice style and emotions) as playing a crucial effect in interaction, which tends to be more immersive experiences in realistic scenarios. Towards this goal, we propose OmniCharacter, a first seamless speech-language personality interaction model to achieve immersive RPAs with low latency. Specifically, OmniCharacter enables agents to consistently exhibit role-specific personality traits and vocal traits throughout the interaction, enabling a mixture of speech and language responses. To align the model with speech-language scenarios, we construct a dataset named OmniCharacter-10K, which involves more distinctive characters (20), richly contextualized multi-round dialogue (10K), and dynamic speech response (135K). Experimental results showcase that our method yields better responses in terms of both content and style compared to existing RPAs and mainstream speech-language models, with a response latency as low as 289ms.
+
+# 1 Introduction
+
+The rapid advancement of Large language models (LLMs) (Wang et al., 2024b; Park et al., 2023; OpenAI, 2022; Li et al., 2022) has demonstrated strong potential in interactive AI, enabling more natural and engaging human-computer interactions (Luo et al., 2024b; Zhang et al., 2023b). One of the
+
+![](images/b6c45a7fce73beeae8857896bc2f8589f355acdf6ffa2ab9d886feeb3083c33d.jpg)  
+Figure 1: (a) Existing role-playing agents (RPAs) concentrate on engaging dialogue in textual form, whereas the role's voice traits are often ignored. (b) Our method considers the importance of the characters' vocal persona, e.g., voice style and emotions, delivering a more seamless and immersive speech-language interaction.
+
+most promising and widely adopted research directions is the development of role-playing agents (RPAs) (Wang et al., 2024a; Tu et al., 2024; Shao et al., 2023), which have been applied in various domains, such as virtual assistants, AI-driven storytelling, and intelligent non-player characters (NPCs) in video games. The focus of RPAs allows LLMs to simulate diverse personas and mimic human-like behavior via predefined role profiles.
+
+Despite significant advances in this field, most existing studies primarily concentrate on replicating dialogues solely in textual form, while often overlooking the role's crucial voice traits such as tone, style, emotions, and pauses. The voice traits are important paralinguistic concepts that deeply reveal individual differences in conversations, enabling a more immersive and personalized roleplaying interaction. As shown in Figure 1 (a), for the same user query, although both agents pro
+
+vide reasonable responses, their unique tones, i.e., Agent A's serious yet calmed and Agent B's angry and commanding, are difficult to convey through text alone, limiting their role-specific expression. Therefore, incorporating personalized voice traits into the current RPAs can help create more realistic and immersive interactions.
+
+To bridge this gap, we introduce OmniCharacter, a first step in realizing speech-language collaborative RPAs for immersive and seamless interaction. Specifically, a Speech-Language Collaborative Model is first built as a base model to perceive both language inputs (i.e., role profile, dialogue contexts, and user text input) and speech input (i.e., user speech input) for semantic alignment. Based on this, we propose a Role Speech Decoder to generate role-specific speech responses containing their unique voice traits. In particular, it mainly consists of two major components: 1) Role-context Guided Speech Token Prediction, which leverages the textual representations from LLM as prior, ensuring the alignment of discrete speech token generation with the corresponding textual contexts. 2) Role-aware Speech Synthesis, which generates waveform containing rich role-related voice traits from speech tokens specific to the given conditions (i.e., profile, dialogue contexts, current query, and speaker embedding). Note that the proposed model can auto-regressively generate text and speech responses with 289ms latency, ensuring a seamless dialogue interaction.
+
+To facilitate the development of OmniCharacter, we further present the OmniCharacter-10K, a speech-language RPAs dataset with detailed profile annotations, diverse dialogues, and vivid audio responses tailored to each role's characteristic. This dataset presents several appealing properties: 1) Large Vocabulary: It includes a total of 20 characters along with 10,072 multi-turn dialogues, along with 135K audio responses. 2) Rich Annotations: Each character is accompanied by a rich profile, highlighting aspects such as personality, voice style, relationships, and past experiences. Further, each dialogue turn is annotated with corresponding speech responses for both users and characters via TTS models (Du et al., 2024; Kim et al., 2021). 3) Dynamic Curation: The dialogue data is generated through interactive conversations between two chatbots, guided by the predefined profiles. The main contributions are as follows:
+
+- We introduce OmniCharacter, a pioneering
+
+step in realizing speech-language collaborative RPAs.
+
+- We construct OmniCharacter-10K, a multi-turn dialogue dataset with detailed role profiles, conversion data, and high-quality audio annotations for advancing RPAs development.  
+- Experimental results demonstrate that our method achieves superior performance on the role-playing benchmark, i.e., CharacterEval and SocialBench as well as providing robust results on general speech benchmarks, i.e., LibriSpeech, AISHELL-2, and LibriTTS.
+
+# 2 Methodology
+
+In this section, we present an in-depth presentation of OmniCharacter, the first RPAs that perceive both speech and language understanding while generating character-specific audio-text responses with low latency. The framework is illustrated in Figure 2.
+
+# 2.1 Overview
+
+The traditional RPAs are designed to make LLMs interact with users or other agents by emulating specific characters. To achieve this, RPAs utilize a profile denoted as  $P$ , along with the ongoing dialogue contexts  $C_n = [q_1, r_1, \dots, q_n]$  to generate a response  $r_n$  consistently with character sets:
+
+$$
+r _ {n} = \operatorname {R P A s} \left(C _ {n}, P\right), \tag {1}
+$$
+
+where  $q_{i}$  represent the  $i$ -th input query and  $r_{i}$  indicates the corresponding response. However, this paradigm relies on LLMs' behavior cloning ability and focuses mainly on textual form, neglecting essential vocal traits like voice style, tone, and emotions of different characters.
+
+To bridge this gap, we introduce OmniCharacter, the first model considering the role-related vocal persona by collaborating speech and language for immersive interaction. Specifically, OmniCharacter mainly contains two components: (1) a Speech-Language Collaborative model to align semantics by processing both language inputs (role profile, dialogue contexts, and text query) and speech input, (2) a Role Speech Decoder that aims to generate role-specific speech responses that reflect their unique voice traits. Thereby, the dialogues contexts  $C_n$  can be redefined as  $U_n = [X_1, Y_1, \ldots, X_n]$ , where  $X_i \in \{X_i^S, X_i^T, X_i^{S + T}\}$  represents the  $i$ -th input query, supporting three types: audio-only
+
+![](images/3e58382acf34bdb6604adddb4f4a0862f072b31c8ca3bb41159511c4ec1f1f0f.jpg)  
+Figure 2: The overview of our OmniCharacter framework. We first build a speech-language collaborative model, a large language model that receives both speech and language inputs for unified modeling. Furthermore, we propose a role speech decoder to synthesize speech responses containing vocal traits of different characters by devising two innovative modules: i) Role-context Guided Speech Token Prediction, which aims to enhance the generation of speech tokens by leveraging the textual representations of the base model as a contextual prior. ii) Role-aware Speech Synthesis, which generates speech responses from speech tokens maintaining unique character personality and voice traits based on several conditions.
+
+$(X_{i}^{S})$ , text-only  $(X_{i}^{T})$ , or audio-text  $(X_{i}^{S + T})$ , resulting in flexible interaction during dialogue. Similarly,  $Y_{n}\in \{Y_{n}^{S},Y_{n}^{T}\}$  indicates the response, owning of two types, i.e., audio response  $(Y_{n}^{S})$  and text response  $(Y_{n}^{T})$ , which maintains the consistent and distinct personality of characters. Next, we elaborate on the details of two components in the subsequent subsections.
+
+# 2.2 Speech-Language Collaborative Model
+
+To fully align speech and language semantics, we first devise a unified speech-language collaborative model. Specifically, it comprises a speech encoder, a speech adaptor, and an LLM.
+
+Speech Encoder. To tackle user or character speech input, we adopt a speech encoder to encode the speech input  $X_{n}^{S}$  of the  $n$ -th dialogue turn to speech sequence  $M_{n}^{S} = [M_{n,1}^{S},\dots,M_{n,N_{f}}^{S}]$ , where  $N_{f}$  indicate the number of audio frames.
+
+Speech Adaptor. To address the high temporal redundancy of the encoded speech sequence above, a down-sampling strategy is further applied by grouping every  $k$  consecutive frame into compact sequence  $Z_{n}^{S} = [Z_{n,1}^{S},\dots,Z_{n,\lfloor N_{s} / k\rfloor}^{S}]$ , where
+
+$$
+Z _ {n, i} ^ {S} = M _ {n, k * i} ^ {S} \oplus M _ {n, k * i + 1} ^ {S} \oplus \dots \oplus M _ {n, k * i + k - 1} ^ {S}. \tag {2}
+$$
+
+This sequence is then encoded by a speech adapter
+
+$\tau$  to make it compatible with the LLM embedding:
+
+$$
+E _ {n} ^ {S} = \tau \left(Z _ {n} ^ {S}\right). \tag {3}
+$$
+
+Large Language Model. We utilize Qwen2.5-7B-Instruct (Yang et al., 2024) as the LLM, a state-of-the-art large language model renowned for its advanced reasoning capabilities to align text and audio sequences for unified modeling. In particular, we process the text inputs, including the role profile  $P$ , dialogue contexts  $U_{n}$ , and the text query  $X_{n}^{T}$  to form an overall sequence  $E_{n}^{T} = \{E_{n,i}^{T}\}_{i=1}^{N_{t}}$  via a text encoder. This sequence is then concatenated with the speech sequence  $E_{n}^{S} = \{E_{n,i}^{S}\}_{i=1}^{N_{s}}$  and sent into LLM. The  $N_{t}$  and  $N_{s}$  indicate the length of text and speech sequences respectively. The prompt template for organizing the overall inputs is showcased in the Appendix A.2. Finally, the LLM is trained via an auto-regressive manner:
+
+$$
+\mathcal {L} _ {\text {l a n g u a g e}} = - \sum_ {i = 1} ^ {N _ {t}} \log P \left(O _ {i} ^ {T} \mid O _ {<   i} ^ {T}\right), \tag {4}
+$$
+
+where  $O_{i}^{T}$  represents the textual token output from the LLM.
+
+# 2.3 Role Speech Decoder
+
+Role-context Guided Speech Token Prediction.
+
+Intuitively, it is straightforward to use the LLM
+
+to predict both text tokens  $O^T$  and speech tokens  $O^S$ . However, we find this strategy makes the model difficult to train, causing duplicate speech token outputs and hallucinations that are inconsistent with the semantics of text tokens. To facilitate stable and accurate speech token prediction, we propose an innovative module known as role-context guided speech token prediction built upon the speech-language collaborative model. Specifically, it employs a lightweight language model (denoted as SpeechLLM) that effectively utilizes the context representations  $H = [H_1, \dots, H_{N_s + N_t}]$  output from LLM to generate speech tokens:
+
+$$
+O ^ {S} = \operatorname {S p e e c h L L M} \left(O _ {m} ^ {S} \mid O _ {1: m - 1} ^ {S}, \phi (H)\right), \tag {5}
+$$
+
+where  $\phi$  is a linear projection layer to map the context representations into the SpeechLLM space. We find that this design not only stabilized the training but also ensured that the generated speech tokens are highly aligned with the contextual semantics, effectively preventing the speech token repetitive output problem. Note that the SpeechLLM also follows an auto-regressive training objective:
+
+$$
+\mathcal {L} _ {s p e e c h} = - \sum_ {i = 1} ^ {N _ {s}} \log P (O _ {i} ^ {S} \mid O _ {<   i} ^ {S}). \qquad (6)
+$$
+
+Role-aware Speech Synthesis. A high-quality speech response that reflects the character's voice traits and persona is important for immersive RPAs. To achieve this, we propose a role-aware speech synthesis module, which generates the audio response containing voice traits of characters from the speech tokens. Following (Zeng et al., 2024), we first decode the speech tokens into the Mel spectrogram and then synthesize the waveform with the generated Mel spectrogram as input. To achieve this, we adopt a conditional flow matching (CFM) model to sample the Mel spectrogram specified by the context representations  $H$ , speech tokens  $O^S$ , and speaker embedding  $v$  (Wang et al., 2023b) extracted from character voice as conditions. The sampling process can be defined by a time-dependent vector field  $T = \{T_1, \dots, T_n\}$ . To improve efficiency, the optimal-transport (OT) flow is used. We summarize the above process as:
+
+$$
+Y _ {n} ^ {S} = \operatorname {O T - C F M} \left(p _ {0} \mid H, v, O ^ {S}\right), \tag {7}
+$$
+
+where  $p_0$  is the initial Mel spectrogram. Ultimately, after obtaining the generated Mel spectrogram, a
+
+HiFi-GAN vocoder serves to synthesize a continuous waveform from the Mel spectrogram, facilitating the conversion of spectral representations into high-fidelity audio signals  $Y_{n}^{S}$  containing characters' voice traits such as voice style and emotions.
+
+# 2.4 Training Strategy
+
+Following (Fang et al., 2024), we adopt a two-stage training strategy for OmniCharacter. In the first stage, we train the speech adapter and LLM to generate text responses based on the text and speech inputs using the objective  $\mathcal{L}_{\text {language }}$  in Equation (4). In the second stage, we only train the linear layer and speech LLM in role-context guided speech token prediction module for speech token prediction by using the objective  $\mathcal{L}_{\text {speech }}$  in Equation (6). For role-aware speech synthesis, we utilize the pretrained conditional flow matching and HiFi-GAN weights from GLM-4-Voice (Zeng et al., 2024) and fine-tune them on high-quality role speech data<sup>2</sup>.
+
+# 3 The OmniCharacter-10K Dataset
+
+To facilitate the model with speech-language scenarios, we construct OmniCharacter-10K, a multi-turn dialogue dataset with distinctive characters with rich and expressive text and speech annotations. In this section, we describe the data collection, annotation, and verification of OmniCharacter-10K. We also introduce the statistics and distribution of it. The illustration of OmniCharacter-10K is showcased in Figure 3.
+
+# 3.1 Data Collection, Annotation, and Verification
+
+Step-1: Character Profile Creation. To make sure the collected characters have distinct traits as well as contain high-quality audio, we select from games since (1) the availability of rich information (e.g., personality, background, and vocal tone), and (2) clear, noise-free character audio recorded by professional voice actors. We end up with 20 foundational characters (10 Chinese and 10 English) from Genshin Impact to construct our dataset. Subsequently, we use a large language model to summarize the meta information for each character, and then instruct this model to expand the information to ensure the diversity and complexity. At last, all character profiles are subjected to rigorous human checks to ensure accuracy and reliability. Moreover, given the diversity of humans
+
+![](images/6eb4edb339e637ec557bda721d0994ba9cce018fb72bca90392d1e84b2a3ba38.jpg)  
+(a) Dataset Construction Pipeline
+
+![](images/123b4847eff393cf2f91ce414866db7caefad5294c45cbb6b048483c6b620969.jpg)  
+Figure 3: Illustration of OmniCharacter-10K. (a) Dataset Construction Pipeline, which consists of four steps: (1) Characters Profile Creation, (2) Dialogue Generation, (3) Speech Synthesis, and (4) Quality Verification. (b) Dataset Properties, which has three appealing properties of our dataset: (1) Large Vocabulary, the dataset includes 20 characters, 10K multi-turn dialogues, and 135K audio responses, (2) Rich Annotations, each character is equipped with rich text and speech annotations, and (3) Dynamic Curation, we generate dialogue data through chatbot interactions based on predefined character profiles.
+
+![](images/a7a0102a337e059b1df75d90632198a3fb9ed9c2811d13f5b3f05515c012f68d.jpg)  
+(b) Dataset Properties
+
+<table><tr><td>Splits</td><td># Characters</td><td>Avg. Turns/Conv.</td><td># Samples</td><td># Speech Hours. (user/character)</td></tr><tr><td>Training</td><td>20</td><td>15</td><td>9,672</td><td>360.3 (176.4/183.9)</td></tr><tr><td>Test</td><td>20</td><td>15</td><td>400</td><td>14.84 (7.15/7.69)</td></tr></table>
+
+Table 1: The statistic of OmniCharacter-10K dataset.
+
+and the open-domain nature of the question, we do not consider the profile for the 'user' side.
+
+Step-2: Dialogue Generation. After selecting the characters and defining their profiles, we generate multi-turn dialogues to evaluate role-playing consistency and conversational quality. Specifically, we deploy two large language models in a simulated interaction: one assumes the role of the target character, while the other acts as a user or a different character. Each model is conditioned on its respective profile, which includes key personality traits, background information, emotions, etc. The dialogue is guided by prompts that encourage natural and persona-consistent responses across multiple turns. This setup enables us to systematically assess how well the models embody their assigned roles, maintain character coherence, and engage in meaningful, contextually appropriate exchanges.
+
+Step-3: Speech Synthesis. The generated multi-turn dialogues consist of open-domain text without corresponding audio annotations. To handle this, we first collected 40K high-quality audio samples for 20 characters, which are professionally recorded by voice actors and extracted from the game's asset files. Then, we utilize these audios as seed data and train a multi-speaker text-to-speech
+
+![](images/8e7aa1ff310e4848fb38cefebfaa696b3fcaa152e9b4df67836c958081e5ebc6.jpg)  
+Figure 4: Left: Distribution of dialogue turns across samples in the OmniCharacter-10K dataset. Right: Distribution of audio duration for user speech and character speech respectively.
+
+![](images/c35e41e0c7ce3d7a05a776fc85c4039183a2f0e68235fc2aabb2c98f68f47021.jpg)
+
+(TTS) model, i.e., VITS (Kim et al., 2021) for the audio synthesis of open-text in dialogue. For the user side, we use the advanced TTS model CosyVoice (Du et al., 2024) to synthesize audio. Each dialogue sample has a  $50\%$  ratio of being assigned a generic male or female voice, ensuring its balance and diverse distribution.
+
+Step-4: Quality Verification. To ensure the quality of the constructed dataset, we manually filter samples based on several criteria. For text data, we performed data filtering by retaining the samples following ABAB (user-role or role1-role2) pattern. Besides, we preserved dialogues longer than three turns, filtering out shorter ones. Furthermore, we remove data that mimics the style of general AI assistance such as 'I am a helpful AI assistant...', as well as unnecessary explanatory prefixes and postfaces. For speech data, we first use the Whisper-largev3 (Radford et al., 2023) to convert the synthesized audio into text and compute the WER metrics with
+
+<table><tr><td rowspan="2">Models</td><td colspan="6">Character Consistency</td><td colspan="4">Conversational Ability</td><td colspan="5">Role-playing Attractiveness</td><td rowspan="2">Avg.</td></tr><tr><td>KE</td><td>KA</td><td>KH</td><td>PB</td><td>PU</td><td>Avg.</td><td>Flu.</td><td>Coh.</td><td>Cons.</td><td>Avg.</td><td>HL</td><td>CS</td><td>ED</td><td>Emp.</td><td>Avg.</td></tr><tr><td colspan="17">Proprietary Models</td></tr><tr><td>BC-NPC-Turbo</td><td>1.802</td><td>2.964</td><td>2.993</td><td>2.910</td><td>3.151</td><td>2.764</td><td>3.578</td><td>3.898</td><td>3.916</td><td>3.798</td><td>3.836</td><td>2.643</td><td>2.336</td><td>2.971</td><td>2.946</td><td>3.169</td></tr><tr><td>MiniMax</td><td>1.835</td><td>2.910</td><td>2.944</td><td>2.774</td><td>3.125</td><td>2.718</td><td>3.609</td><td>3.932</td><td>3.811</td><td>3.784</td><td>3.768</td><td>2.672</td><td>2.150</td><td>3.017</td><td>2.902</td><td>3.134</td></tr><tr><td>GPT-3.5</td><td>1.716</td><td>2.339</td><td>2.212</td><td>1.921</td><td>2.316</td><td>2.101</td><td>2.629</td><td>2.917</td><td>2.700</td><td>2.749</td><td>2.565</td><td>2.422</td><td>1.660</td><td>2.526</td><td>2.293</td><td>2.381</td></tr><tr><td>GPT-4</td><td>2.250</td><td>2.855</td><td>2.785</td><td>2.721</td><td>2.873</td><td>2.697</td><td>3.332</td><td>3.669</td><td>3.343</td><td>3.448</td><td>3.143</td><td>3.184</td><td>2.153</td><td>3.010</td><td>2.873</td><td>3.006</td></tr><tr><td colspan="17">Open-sourced Models</td></tr><tr><td>ChatGLM3-6B</td><td>2.016</td><td>2.792</td><td>2.704</td><td>2.455</td><td>2.812</td><td>2.556</td><td>3.269</td><td>3.647</td><td>3.283</td><td>3.399</td><td>3.064</td><td>2.932</td><td>1.969</td><td>2.993</td><td>2.739</td><td>2.898</td></tr><tr><td>Baichuan2-7B</td><td>1.813</td><td>2.849</td><td>2.929</td><td>2.830</td><td>3.081</td><td>2.700</td><td>3.551</td><td>3.894</td><td>3.827</td><td>3.757</td><td>3.670</td><td>2.728</td><td>2.115</td><td>2.984</td><td>2.874</td><td>3.110</td></tr><tr><td>Baichuan2-13B</td><td>1.802</td><td>2.869</td><td>2.946</td><td>2.808</td><td>3.081</td><td>2.701</td><td>3.596</td><td>3.924</td><td>3.864</td><td>3.759</td><td>3.700</td><td>2.703</td><td>2.136</td><td>3.021</td><td>2.890</td><td>3.116</td></tr><tr><td>InternLM-7B</td><td>1.782</td><td>2.800</td><td>2.781</td><td>2.719</td><td>3.016</td><td>2.620</td><td>3.527</td><td>3.823</td><td>3.744</td><td>3.698</td><td>3.546</td><td>2.622</td><td>2.070</td><td>2.897</td><td>2.784</td><td>2.983</td></tr><tr><td>InternLM-20B</td><td>1.945</td><td>2.916</td><td>2.920</td><td>2.753</td><td>3.041</td><td>2.715</td><td>3.576</td><td>3.943</td><td>3.717</td><td>3.745</td><td>3.582</td><td>2.885</td><td>2.132</td><td>3.047</td><td>2.911</td><td>3.123</td></tr><tr><td>CharacterGLM</td><td>1.640</td><td>2.819</td><td>2.738</td><td>2.301</td><td>2.969</td><td>2.493</td><td>3.414</td><td>3.717</td><td>3.737</td><td>3.623</td><td>3.738</td><td>2.265</td><td>1.966</td><td>2.812</td><td>2.695</td><td>2.937</td></tr><tr><td>Llama-3.1-8B</td><td>2.197</td><td>2.701</td><td>2.615</td><td>3.130</td><td>2.704</td><td>2.669</td><td>3.059</td><td>3.477</td><td>3.071</td><td>3.202</td><td>2.922</td><td>2.934</td><td>2.634</td><td>2.759</td><td>2.812</td><td>2.894</td></tr><tr><td>Qwen-7B</td><td>1.956</td><td>2.728</td><td>2.633</td><td>2.605</td><td>2.780</td><td>2.540</td><td>3.187</td><td>3.564</td><td>3.229</td><td>3.327</td><td>3.036</td><td>2.791</td><td>2.052</td><td>2.838</td><td>2.679</td><td>2.848</td></tr><tr><td>Qwen-14B</td><td>1.988</td><td>2.800</td><td>2.811</td><td>2.744</td><td>2.900</td><td>2.649</td><td>3.351</td><td>3.765</td><td>3.510</td><td>3.542</td><td>3.354</td><td>2.871</td><td>2.237</td><td>2.970</td><td>2.858</td><td>3.016</td></tr><tr><td>Qwen2-7B-Instruct</td><td>1.966</td><td>2.537</td><td>2.412</td><td>2.313</td><td>2.436</td><td>2.333</td><td>2.864</td><td>3.171</td><td>2.743</td><td>2.926</td><td>2.655</td><td>2.612</td><td>1.867</td><td>2.654</td><td>2.447</td><td>2.569</td></tr><tr><td>Qwen2.5-7B-Instruct</td><td>2.207</td><td>2.740</td><td>2.633</td><td>2.700</td><td>2.614</td><td>2.579</td><td>3.125</td><td>3.401</td><td>2.971</td><td>3.166</td><td>2.786</td><td>2.871</td><td>2.180</td><td>2.826</td><td>2.666</td><td>2.804</td></tr><tr><td>Qwen2.5-7B-Instruct†</td><td>2.172</td><td>3.012</td><td>2.887</td><td>3.526</td><td>2.879</td><td>2.895</td><td>3.321</td><td>3.657</td><td>3.326</td><td>3.434</td><td>3.278</td><td>3.127</td><td>2.728</td><td>3.217</td><td>3.087</td><td>3.138</td></tr><tr><td>OmniCharacter w/o audio</td><td>2.188</td><td>3.024</td><td>2.926</td><td>3.547</td><td>2.955</td><td>2.928</td><td>3.402</td><td>3.732</td><td>3.432</td><td>3.522</td><td>3.303</td><td>3.266</td><td>2.933</td><td>3.172</td><td>3.168</td><td>3.206</td></tr><tr><td>OmniCharacter</td><td>2.230</td><td>3.040</td><td>2.918</td><td>3.531</td><td>2.988</td><td>2.941</td><td>3.369</td><td>3.768</td><td>3.410</td><td>3.516</td><td>3.374</td><td>3.261</td><td>3.002</td><td>3.187</td><td>3.206</td><td>3.221</td></tr></table>
+
+Table 2: Performance comparison with state-of-the-art methods on CharacterEval. We evaluate the models across three dimensions including: (A) Character Consistency: (a-1) KE: Knowledge-Exposure, (a-2) KA: Knowledge-Accuracy, (a-3) KH: Knowledge-Hallucination, (a-4) PB: Persona-Behavior, (a-5) PU: Persona-Utterance. (B) Conversational Ability: (b-1) Flu.: Fluency, (b-2) Coh.: Coherency, (b-3) Cons.: Consistency. (C) Role-playing Attractiveness: (c-1) HL: Human-Likeness, (c-2) CS: Communication Skill, (c-3) ED: Expression Diversity, (c-4) Emp.: Empathy.  $\dagger$  means we finetune pretrained Qwen-2.5-7B-Instruct on the textual portion of OmniCharacter-10K.
+
+the corresponding text, where the WER greater than 10 are rejected. Then, we use WavLLM (Hu et al., 2024) to compute the similarity between the synthesized audio and reference audio, where the similarity less than 0.8 threshold is also rejected.
+
+# 3.2 Statistics of OmniCharacter-10K
+
+Table 1 presents the statistics of the OmniCharacter-10K dataset, which includes a total of 20 characters with 9,672 training samples and 400 test samples. The training and test sets are rich in speech annotations, with 360.3 hours and 14.84 hours of audio data, respectively. In addition, Figure 4 shows the distribution of dialogue turns (left) and audio duration (right) across the dataset. On one hand, more than  $80\%$  of the dialogues are longer than 10 turns, highlighting the dataset's focus on multi-turn interactions. On the other hand, the audio duration statistics further reveal that the speech in the dataset is sufficiently long to enable immersive dialogue experiences. These statistics emphasize the dataset's capability to support the evaluation of RPAs in facilitating immersive dialogues.
+
+# 4 Experiments
+
+# 4.1 Experimental Setups
+
+Model Configuration. We employ the Whisper-large-v3 as the speech encoder and Qwen-2.5-7B
+
+Instruct as the LLM in speech-language collaborative model. In role-context guided speech token prediction, we employ the lightweight Qwen2.5-0.5B-Instruct (Team, 2024) model as SpeechLLM to generate speech tokens. To facilitate this process, a speech tokenizer derived from GLM-4-Voice (Zeng et al., 2024) with 16K vocabulary size is utilized to extract discrete speech tokens. We resample all speech data to a frequency of 16Khz. The speech-language collaborative model and role-context guided speech token prediction are initialized using pre-trained parameters from OpenOmni (Luo et al., 2025), and we conduct finetuning on the OmniCharacter-10K train set. The weights of role-aware speech synthesis are initialized from pretrained checkpoint of GLM-4-Voice.
+
+Training Details. The OmniCharacter employs a two-stage training strategy as described in Section 2.4. In the first stage, we utilize the AdamW optimizer with a warmup ratio of 0.3, gradually increasing the learning rate until it reaches 5e-4. The second stage retains the parameters established in the first stage, and we only adjusted the final learning rate to 5e-5. We set the batch size as 32 for both two stages. The entire training process is finished within 3 epochs on  $8 \times \mathrm{A}100$  GPUs.
+
+Datasets. The evaluation leverages the following datasets: CharacterEval (Tu et al., 2024) and Social-
+
+<table><tr><td rowspan="2">Models</td><td colspan="5">Individual Level</td><td colspan="4">Group Level</td><td rowspan="2">Avg.</td></tr><tr><td>SA Style</td><td>SA Know.</td><td>EP Situ.</td><td>EP Emo.</td><td>CM Short</td><td>CM Long</td><td>Pos.</td><td>Neu.</td><td>Neg.</td></tr><tr><td>Qwen-72B-Chat</td><td>83.87</td><td>90.64</td><td>53.10</td><td>52.89</td><td>83.29</td><td>73.15</td><td>91.53</td><td>73.44</td><td>63.82</td><td>73.97</td></tr><tr><td>LLaMA-2-70B-Chat</td><td>67.61</td><td>70.78</td><td>35.74</td><td>38.47</td><td>45.57</td><td>26.74</td><td>69.87</td><td>45.29</td><td>39.37</td><td>48.83</td></tr><tr><td>Qwen-14B-Chat</td><td>77.06</td><td>86.15</td><td>45.71</td><td>43.78</td><td>65.32</td><td>51.37</td><td>78.32</td><td>58.25</td><td>59.21</td><td>62.80</td></tr><tr><td>LLaMA-2-13B-Chat</td><td>57.62</td><td>65.51</td><td>37.12</td><td>32.56</td><td>30.43</td><td>29.82</td><td>66.38</td><td>42.25</td><td>26.27</td><td>43.11</td></tr><tr><td>LLaMA-2-7B-Chat</td><td>48.76</td><td>51.23</td><td>31.23</td><td>28.91</td><td>25.38</td><td>21.89</td><td>44.98</td><td>24.19</td><td>27.67</td><td>33.80</td></tr><tr><td>Mistral-7B</td><td>50.12</td><td>61.17</td><td>36.48</td><td>31.72</td><td>31.78</td><td>25.42</td><td>65.67</td><td>46.34</td><td>28.96</td><td>41.96</td></tr><tr><td>Qwen-7B-Chat</td><td>66.44</td><td>71.16</td><td>41.68</td><td>40.68</td><td>67.45</td><td>53.45</td><td>75.61</td><td>52.78</td><td>43.11</td><td>56.93</td></tr><tr><td>OmniCharacter</td><td>70.93</td><td>65.68</td><td>58.03</td><td>38.24</td><td>59.76</td><td>57.86</td><td>69.11</td><td>63.26</td><td>43.98</td><td>58.53</td></tr></table>
+
+Table 3: Performance comparison with state-of-the-art methods on SocialBench. We evaluate the models across several dimensions including: (1) Self-Awareness: role style (SA Style) and role knowledge (SA Know.) (2) Emotional Perception: situation understanding (EP Situ.) and emotion detection (EP Emo.) (3) Conversation Memory: short-term (CM Short) and long-term (CM Long) (4) Social Preference: positive (Pos.), neutral (Neu.), and negative (Neg.).  
+
+<table><tr><td rowspan="2">Models</td><td colspan="2">S2TIF</td><td colspan="2">S2SIF</td></tr><tr><td>Content ↑</td><td>Style ↑</td><td>Content ↑</td><td>Style ↑</td></tr><tr><td>SpeechGPT</td><td>1.25</td><td>1.28</td><td>1.79</td><td>1.85</td></tr><tr><td>LLaMA-Omni</td><td>2.46</td><td>2.71</td><td>2.26</td><td>2.38</td></tr><tr><td>OmniCharacter</td><td>3.89</td><td>4.18</td><td>2.64</td><td>2.55</td></tr></table>
+
+Table 4: Performance comparison with state-of-the-art methods on OmniCharacter-10K test split. ChatGPT scores for S2TIF, S2SIF tasks, and alignment scores between speech and text responses.
+
+Bench (Chen et al., 2024) are used to testify models' basic language understanding. OmniCharacter-10K test set is employed for character-aware speech-language evaluation to test speech generation and alignment abilities. LibriSpeech (Panayotov et al., 2015), AISHELL-2 (Du et al., 2018), and LibriTTS (Zen et al., 2019) are general speech datasets for model's generalization test.
+
+# 4.2 Performance Comparison
+
+Basic Language Understanding. To evaluate the effectiveness of our OmniCharacter in basic language understanding, we compare our method with RPAs methods on CharacterEval (Tu et al., 2024) benchmark as shown in Table 2. Note that the characters tested in CharacterEval differ from those in our training data. The results illustrate that our model outperforms the proprietary and open-source models on most metrics. Particularly, compared with models specifically designed for role-playing task including BC-NPC-Turbo and MiniMax, our method achieves considerable improvements. Moreover, OmniCharacter surpasses open-sourced models with larger scales such as InterLM-20B and Qwen-14B, achieving superior
+
+performance in both character consistency and role-playing attractiveness. Further, we evaluate our method on SocialBench (Chen et al., 2024) as shown in Table 3. As observed, our method improves performance across most metrics when compared to the same-size models. Although OmniCharacter is not as optimal as larger-scale models, it still delivers comparable performance, securing the second-best position in terms of the average score. The results presented above provide compelling evidence that the integration of audio modality enhances basic language understanding.
+
+Character-aware Speech-Language Evaluation. To thoroughly evaluate the speech-language collaboration ability of OmniCharacter for immersive and personalized role-playing interaction, we conduct evaluations from two perspectives: metrics-based and human-based evaluations.
+
+The metrics-based evaluation consists of three tasks (Fang et al., 2024): (1) speech-to-text instruction-following (S2TIF): whether the model's text response correctly answers the input speech query; (2) speech-to-speech instruction-following (S2SIF): whether the model's speech response correctly answers the input speech query. We ask GPT-4o to score the response in a range from 1 to 5 from content and style aspects. As shown in Table 4, our model consistently enhances performance across all metrics, achieving superior results compared to LLaMA-Omni and SpeechGPT. These results confirm that OmniCharacter is capable of effectively handling user or characters' instructions, while simultaneously generating consistent text and speech responses.
+
+In terms of human-based evaluation, five experts are assigned to examine the model's outputs across
+
+<table><tr><td>Models</td><td>Flu.</td><td>Cons.</td><td>Emo.</td><td>Cla.</td><td>App.</td><td>Imm.</td></tr><tr><td>SpeechGPT</td><td>6.12</td><td>3.86</td><td>3.75</td><td>6.89</td><td>4.23</td><td>3.64</td></tr><tr><td>LLaMA-Omni</td><td>6.88</td><td>4.27</td><td>3.44</td><td>6.69</td><td>4.78</td><td>4.68</td></tr><tr><td>OmniCharacter</td><td>7.97</td><td>6.84</td><td>6.23</td><td>7.88</td><td>5.63</td><td>8.52</td></tr></table>
+
+six voice-related dimensions as shown in Table 5. Each expert is asked to score the responses on a scale from 1 to 10 for each dimension. From the table, we figure out that our method reaches the best performance across all dimensions, proving the superiority of our model which is capable of providing immersive interactions for users.
+
+# Generalization on General Speech Benchmarks.
+
+To better prove the robust ability of OmniCharacter, we evaluated the performance of our model on three widely used speech datasets, LibriSpeech, AISHell-2, and LibriTTS, focusing on both Automatic Speech Recognition (ASR) and Text-to-Speech (TTS) tasks. As presented in Table 6, our model achieves comparable performance compared with previous models, demonstrating its generalization ability. These results further highlight the robustness of OmniCharacter in generating coherent and accurate outputs, showcasing its potential in real-world speech applications.
+
+# 4.3 Impact of Audio Modality
+
+To assess the impact of the audio modality on roleplaying agents (RPAs), we conduct an ablation study by removing audio input from OmniCharacter and evaluating the model on the CharacterEval benchmark, as shown in Table 7. We report average scores across all metrics for clarity. The results indicate that incorporating audio significantly improves character consistency and conversational ability, while achieving comparable performance in roleplaying attractiveness. These findings highlight the importance of audio in enhancing language understanding and contributing to more natural, expressive role-playing interactions.
+
+# 4.4 Similarity of Synthesized Voice for Characters
+
+To verify the model's capability of synthesizing character-specific speech containing their voice traits, we use WavLLM to compute the cosine
+
+Table 5: Performance comparison with state-of-the-art methods on OmniCharacter-10K test split. Human score (averaged from 5 experts) across six voice-related dimensions: fluency (Flu.), consistency (Cons.), emotional expression (Emo.), clarity (Cla), appropriateness (App.), and immersion (Imm.).  
+
+<table><tr><td rowspan="2">Models</td><td colspan="2">LibriSpeech (ASR-WER)</td><td>AISHELL-2 (ASR-CER)</td><td>LibriTTS (TTS-WER)</td></tr><tr><td>test-clean</td><td>test-other</td><td>test</td><td>test-clean</td></tr><tr><td>CosyVoice 1.0</td><td>-</td><td>-</td><td>-</td><td>3.17</td></tr><tr><td>Whisper-large-v3</td><td>2.50</td><td>4.53</td><td>-</td><td>-</td></tr><tr><td>Qwen2-Audio</td><td>2.00</td><td>4.50</td><td>3.30</td><td>-</td></tr><tr><td>OmniCharacter</td><td>3.26</td><td>4.23</td><td>6.62</td><td>7.23</td></tr></table>
+
+Table 6: Performance comparison with the state-of-art methods on general speech benchmarks including LibriSpeech, AISHELL-2, and LibriTTS. The model is evaluated on two tasks, i.e., Automatic Speech Recognition (ASR) and Text-to-Speech (TTS). WER: Word Error Rate. CER: Character Error Rate.  
+
+<table><tr><td>Models</td><td>Character Consistency</td><td>Conversational Ability</td><td>Role-playing Attractiveness</td></tr><tr><td>OmniCharacter w/o audio</td><td>2.928</td><td>3.522</td><td>3.168</td></tr><tr><td>OmniCharacter</td><td>2.941</td><td>3.516</td><td>3.206</td></tr></table>
+
+Table 7: Ablation study of model training with or without audio modality. For simplicity, we only report the average score across three dimensions on CharacterEval.
+
+similarity between the reference speech and the generated speech on OmniCharacter-10K test set, as shown in Figure 5. The comparison methods include state-of-the-art TTS systems, i.e., VITS, MaskGCT, CosyVoice1.0, and CosyVoice2.0. For our model, we directly input the corresponding text to generate the speech output. For other models, we leverage their voice cloning capabilities by providing reference speech, along with the given text, to synthesize the corresponding speech. If the cosine similarity between the output speech and the reference speech is greater than 0.9, we classify them as belonging to the same character. As observed, our model consistently achieves higher cosine similarity for both Chinese and English characters' voices. These results highlight the effectiveness of our approach in preserving the voice traits of characters, achieving a more interactive experiment.
+
+# 4.5 Learning Discriminative Character Speech Embeddings
+
+The OmniCharacter generates the character-specific voice by utilizing the speech embedding as one of the conditions. To evaluate the discrimination of speaker embedding between different characters, we compute their cosine similarity as shown in Figure 6. Specifically, we randomly sampled three sets of data of 20 characters for experimental comprehensiveness. We also included the user's speaker embedding for comparison. The results indicate that the similarity values between the speech embeddings of different characters and users are very low, demonstrating their strong discriminabil
+
+<table><tr><td>Models</td><td>GPT-4o</td><td>LLaMA-Omni</td><td>OmniCharacter</td></tr><tr><td>Latency (ms)</td><td>320</td><td>226</td><td>289</td></tr></table>
+
+Table 8: Speech response latency of different models.
+
+![](images/f3e9eded5d4ecdad955c4392be74aec4951b93f867237252c75ab07db0c29275.jpg)  
+Figure 5: Comparison of character voice similarity on OmniCharacter-10K test set. Our model generates audio responses that better match the timbre of characters.
+
+ity. This confirms the robustness of our method in controlling character's voice traits.
+
+# 4.6 Speech Response Latency
+
+To demonstrate the seamless interaction capability of the proposed model, we evaluate the speech response latency of different models, as shown in Table 8. From the table, we observe that our model outperforms the GPT-4o, achieving a lower response latency. It is observed that the latency of our method is higher than LLaMA-Omni. The possible reason is that LLaMA-Omni only supports general female voice, while we introduce additional modules considering voice traits of different roles, thus the increased latency is acceptable. Overall, these results underscore the seamless dialogue ability of our method, demonstrating its effectiveness in real-time speech-language interaction for RPAs.
+
+# 5 Conclusion
+
+In this paper, we present OmniCharacter, the first step toward creating a seamless and immersive RPAs model that integrates both speech and language personality interaction. Unlike existing methods that primarily focus on text-based dialogues, our approach emphasizes the importance of character voice traits, e.g., voice style and emotions, which enable agents to reflect both personality and vocal traits throughout interactions, delivering a harmonious blend of speech and language responses. To facilitate the speech-language scenarios for RPAs, we construct the OmniCharacter-10K dataset, which includes distinct characters, contextually rich multi-turn dialogues, and dynamic speech responses. Experimental results demon
+
+![](images/b53b048ab9132e7815ac749a930534f80e4bd3ea435113416f5b21601efd3285.jpg)  
+Figure 6: Speech embedding discriminability for 20 characters and 4 users. We randomly sampled three sets of data for experimental comprehensiveness. The low similarity values demonstrate strong separation of character voices, enabling precise voice control.
+
+![](images/98d088897274e957540288eeb6246724f9fb62c1a328d55d78b078f88c472463.jpg)
+
+![](images/21a3faf181d5ac79671955c59c9e602147a7b4102e7d364bfeef71dc474316df.jpg)
+
+strate that OmniCharacter outperforms existing RPAs models and general speech-language models in both content fidelity and expressive style, delivering a seamless and immersive user experience.
+
+# 6 Limitations
+
+The proposed OmniCharacter advances immersive role-playing agents (RPAs) by incorporating characters' voice traits through a speech-language personality interaction framework. While effective, several limitations remain. First, performance relies heavily on the quality of the pre-trained language models and speech synthesis systems, which may be affected by their alignment and training data. Second, the current design is limited to two-character dialogues, making it less adaptable to multi-role interactions. Third, although the OmniCharacter-10K dataset includes diverse scenarios, it may not fully capture the richness of real-world conversations. Expanding the dataset is a key direction for improving generalization. Lastly, while the latency is currently low, further optimization is necessary to enhance the system's performance and ensure a more seamless experience for real-time usage.
+
+# Acknowledgments
+
+This study is supported by Alibaba Research Intern Program, grants from the National Natural Science Foundation of China (Grant No.62425208, No. U22A2097, No. U23A20315, No. 62020106008, No. 82441006, No,62402094, No. 62376262), Shenzhen Science and Technology Program (No.JCYJ20240813114208012), the National Key Research and Development Program of China (2022YFF0902100), the Natural Science Foundation of Guangdong Province of China (2025B1515020032, 2024A1515030166), and Sichuan Science and Technology Program (No.2025ZNSFSC1493).
+
+# References
+
+Zalán Borsos, Raphaël Marinier, Damien Vincent, Eugene Kharitonov, Olivier Pietquin, Matt Sharifi, Dominik Roblek, Olivier Teboul, David Grangier, Marco Tagliasacchi, and Neil Zeghidour. 2023. Audiolm: a language modeling approach to audio generation. Preprint, arXiv:2209.03143.  
+Hongzhan Chen, Hehong Chen, Ming Yan, Wenshen Xu, Xing Gao, Weizhou Shen, Xiaojun Quan, Chenliang Li, Ji Zhang, Fei Huang, et al. 2024. Social-bench: Sociality evaluation of role-playing conversational agents. arXiv preprint arXiv:2403.13679.  
+Alexandre Défossez, Laurent Mazaré, Manu Orsini, Amélie Royer, Patrick Pérez, Hervé Jégou, Edouard Grave, and Neil Zeghidour. 2024. Moshi: a speech-text foundation model for real-time dialogue. arXiv preprint arXiv:2410.00037.  
+Jiayu Du, Xingyu Na, Xuechen Liu, and Hui Bu. 2018. Aishell-2: Transforming mandarin asr research into industrial scale. arXiv preprint arXiv:1808.10583.  
+Zhihao Du, Qian Chen, Shiliang Zhang, Kai Hu, Heng Lu, Yexin Yang, Hangrui Hu, Siqi Zheng, Yue Gu, Ziyang Ma, et al. 2024. Cosyvoice: A scalable multilingual zero-shot text-to-speech synthesizer based on supervised semantic tokens. arXiv preprint arXiv:2407.05407.  
+Qingkai Fang, Shoutao Guo, Yan Zhou, Zhengrui Ma, Shaolei Zhang, and Yang Feng. 2024. Llama-omni: Seamless speech interaction with large language models. arXiv preprint arXiv:2409.06666.  
+Shujie Hu, Long Zhou, Shujie Liu, Sanyuan Chen, Lingwei Meng, Hongkun Hao, Jing Pan, Xunying Liu, Jinyu Li, Sunit Sivasankaran, et al. 2024. Wavllm: Towards robust and adaptive speech large language model. arXiv preprint arXiv:2404.00656.  
+Jaehyeon Kim, Jungil Kong, and Juhee Son. 2021. Conditional variational autoencoder with adversarial learning for end-to-end text-to-speech. In ICML, pages 5530-5540. PMLR.  
+Jinlong Li, Zequn Jie, Xu Wang, Xiaolin Wei, and Lin Ma. 2022. Expansion and shrinkage of localization for weakly-supervised semantic segmentation. Advances in neural information processing systems, 35:16037-16051.  
+Jinlong Li, Cristiano Saltori, Fabio Poiesi, and Nicu Sebe. 2025. Cross-modal and uncertainty-aware agglomeration for open-vocabulary 3d scene understanding. arXiv preprint arXiv:2503.16707.  
+Run Luo, Yunshui Li, Longze Chen, Wanwei He, Ting-En Lin, Ziqiang Liu, Lei Zhang,zikai Song, Xiaobo Xia, Tongliang Liu, et al. 2024a. Deem: Diffusion models serve as the eyes of large language models for image perception. arXiv preprint arXiv:2405.15232.
+
+Run Luo, Ting-En Lin, Haonan Zhang, Yuchuan Wu, Xiong Liu, Min Yang, Yongbin Li, Longze Chen, Jiaming Li, Lei Zhang, et al. 2025. Openomni: Large language models pivot zero-shot omnimodal alignment across language with real-time self-aware emotional speech synthesis. arXiv preprint arXiv:2501.04561.  
+Run Luo, Haonan Zhang, Longze Chen, Ting-En Lin, Xiong Liu, Yuchuan Wu, Min Yang, Minzheng Wang, Pengpeng Zeng, Lianli Gao, et al. 2024b. Mmevol: Empowering multimodal large language models with evol-instruct. arXiv preprint arXiv:2409.05840.  
+Openai OpenAI. 2022. Openai: Introducing chatgpt. URL https://openai.com/blog/chatgpt.  
+Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. 2015. Librispeech: an asr corpus based on public domain audio books. In ICASSP, pages 5206-5210. IEEE.  
+Joon Sung Park, Joseph O'Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S Bernstein. 2023. Generative agents: Interactive simulacra of human behavior. In Proceedings of the 36th annual acm symposium on user interface software and technology, pages 1-22.  
+Letian Peng and Jingbo Shang. 2024. Quantifying and optimizing global faithfulness in persona-driven roleplaying. arXiv preprint arXiv:2405.07726.  
+Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. 2023. Robust speech recognition via large-scale weak supervision. In ICML, pages 28492-28518. PMLR.  
+Paul K Rubenstein, Chulayuth Asawaroengchai, Duc Dung Nguyen, Ankur Bapna, Zalán Borsos, Félix de Chaumont Quity, Peter Chen, Dalia El Badawy, Wei Han, Eugene Kharitonov, et al. 2023. Audiopalm: A large language model that can speak and listen. arXiv preprint arXiv:2306.12925.  
+Nafis Sadeq, Zhouhang Xie, Byungkyu Kang, Prarit Lamba, Xiang Gao, and Julian McAuley. 2024. Mitigating hallucination in fictional character role-play. arXiv preprint arXiv:2406.17260.  
+Yunfan Shao, Linyang Li, Junqi Dai, and Xipeng Qiu. 2023. Character-llm: A trainable agent for roleplaying. In EMNLP, pages 13153-13187.  
+Qwen Team. 2024. Qwen2.5: A party of foundation models.  
+Quan Tu, Shilong Fan, Zihang Tian, and Rui Yan. 2024. Charactereval: A chinese benchmark for role-playing conversational agent evaluation. arXiv preprint arXiv:2401.01275.  
+Chengyi Wang, Sanyuan Chen, Yu Wu, Ziqiang Zhang, Long Zhou, Shujie Liu, Zhuo Chen, Yanqing Liu, Huaming Wang, Jinyu Li, et al. 2023a. Neural codec language models are zero-shot text to speech synthesizers. arXiv preprint arXiv:2301.02111.
+
+Hui Wang, Siqi Zheng, Yafeng Chen, Luyao Cheng, and Qian Chen. 2023b. Cam++: A fast and efficient network for speaker verification using context-aware masking.  
+Lei Wang, Jianxun Lian, Yi Huang, Yanqi Dai, Haoxuan Li, Xu Chen, Xing Xie, and Ji-Rong Wen. 2024a. Characterbox: Evaluating the role-playing capabilities of llms in text-based virtual worlds. arXiv preprint arXiv:2412.05631.  
+Lei Wang, Chen Ma, Xueyang Feng, Zeyu Zhang, Hao Yang, Jingsen Zhang, Zhiyuan Chen, Jiakai Tang, Xu Chen, Yankai Lin, et al. 2024b. A survey on large language model based autonomous agents. Frontiers of Computer Science, 18(6):186345.  
+Tianrui Wang, Long Zhou, Ziqiang Zhang, Yu Wu, Shujie Liu, Yashesh Gaur, Zhuo Chen, Jinyu Li, and Furu Wei. 2024c. Viola: Conditional language models for speech recognition, synthesis, and translation. IEEE/ACM Transactions on Audio, Speech, and Language Processing.  
+Xintao Wang, Yunze Xiao, Jen-tse Huang, Siyu Yuan, Rui Xu, Haoran Guo, Quan Tu, Yaying Fei, Ziang Leng, Wei Wang, et al. 2024d. Incharacter: Evaluating personality fidelity in role-playing agents through psychological interviews. In ACL, pages 1840-1873.  
+Zekun Moore Wang, Zhongyuan Peng, Haoran Que, Jiaheng Liu, Wangchunshu Zhou, Yuhan Wu, Hongcheng Guo, Ruitong Gan, Zehao Ni, Jian Yang, et al. 2023c. Rolellm: Benchmarking, eliciting, and enhancing role-playing abilities of large language models. arXiv preprint arXiv:2310.00746.  
+An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. 2024. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115.  
+Tao Yang, Yuhua Zhu, Xiaojun Quan, Cong Liu, and Qifan Wang. 2025. Psyplay: Personality-infused role-playing conversational agents. arXiv preprint arXiv:2502.03821.  
+Yeyong Yu, Runsheng Yu, Haojie Wei, Zhanqiu Zhang, and Quan Qian. 2024. Beyond dialogue: A profile-dialogue alignment framework towards general role-playing language model. arXiv preprint arXiv:2408.10903.  
+Heiga Zen, Viet Dang, Rob Clark, Yu Zhang, Ron J Weiss, Ye Jia, Zhifeng Chen, and Yonghui Wu. 2019. Libritts: A corpus derived from librispeech for text-to-speech. arXiv preprint arXiv:1904.02882.  
+Aohan Zeng, Zhengxiao Du, Mingdao Liu, Kedong Wang, Shengmin Jiang, Lei Zhao, Yuxiao Dong, and Jie Tang. 2024. Glm-4-voice: Towards intelligent and human-like end-to-end spoken chatbot. arXiv preprint arXiv:2412.02612.
+
+Dong Zhang, Shimin Li, Xin Zhang, Jun Zhan, Pengyu Wang, Yaqian Zhou, and Xipeng Qiu. 2023a. Speechgpt: Empowering large language models with intrinsic cross-modal conversational abilities. arXiv preprint arXiv:2305.11000.  
+Dong Zhang, Xin Zhang, Jun Zhan, Shimin Li, Yaqian Zhou, and Xipeng Qiu. 2024a. Speechgpt-gen: Scaling chain-of-information speech generation. arXiv preprint arXiv:2401.13527.  
+Haonan Zhang, Pengpeng Zeng, Lianli Gao, Xinyu Lyu, Jingkuan Song, and Heng Tao Shen. 2023b. Spt: Spatial pyramid transformer for image captioning. TCSVT, 34(6):4829-4842.  
+Qinglin Zhang, Luyao Cheng, Chong Deng, Qian Chen, Wen Wang, Siqi Zheng, Jiaqing Liu, Hai Yu, Chao-hong Tan, Zhihao Du, et al. 2024b. Omniflatten: An end-to-end gpt model for seamless voice conversation. arXiv preprint arXiv:2410.17799.  
+Dong Zhao, Jinlong Li, Shuang Wang, Mengyao Wu, Qi Zang, Nicu Sebe, and Zhun Zhong. 2025. Fishertune: Fisher-guided robust tuning of vision foundation models for domain generalized segmentation. arXiv preprint arXiv:2503.17940.  
+Jinfeng Zhou, Zhuang Chen, Dazhen Wan, Bosi Wen, Yi Song, Jifan Yu, Yongkang Huang, Libiao Peng, Jiaming Yang, Xiyao Xiao, et al. 2023. Characterglm: Customizing chinese conversational ai characters with large language models. arXiv preprint arXiv:2311.16832.
+
+# A Appendix
+
+# A.1 Related Works
+
+Role-Playing Agents. Benefiting from the powerful summarization and imitation capabilities of large language models (LLMs) (Shao et al., 2023; Li et al., 2025; Zhao et al., 2025), role-playing agents (RPAs) have made significant progress in recent years. Prior research primarily focuses on replicating the knowledge, experiences, and intent of characters (Yang et al., 2025; Sadeq et al., 2024; Yu et al., 2024; Zhou et al., 2023). For instance, (Sadeq et al., 2024) reduces hallucination by adjusting parametric knowledge influence based on a pre-calibrated confidence threshold. (Shao et al., 2023) seeks to train an agent with profile and experience perception, replacing limited prompts for LLM instruction. Due to the lack of comprehensive benchmarks, later works focused on building character-specific datasets and new evaluation metrics. (Peng and Shang, 2024; Wang et al., 2024d,a, 2023c). (Wang et al., 2024d) aims to evaluate the personality fidelity of RPAs using psychological scales. (Wang et al., 2024a) devises a simulation sandbox to generate fine-grained character behavior trajectories to evaluate RPAs capabilities. However, most existing studies primarily focus on replicating dialogues in textual form, often overlooking the role's essential voice traits, such as tone, style, emotions, and pauses. To bridge this gap, we propose OmniCharacter, a model that seamlessly combines speech and language to ensure immersive RPAs interactions. Speech-Language Models. Recent advancements in speech-language models have significantly improved human-machine interactions (Fang et al., 2024; Zhang et al., 2023a; Luo et al., 2024a; Zeng et al., 2024; Zhang et al., 2024a). With the advancement of language models in the natural language processing field, early work attempted to enable language models to generate both speech tokens and text tokens via a decoder-only architectural (Wang et al., 2024c; Borsos et al., 2023; Wang et al., 2023a). With the development of large language models (LLMs), more recent studies have integrated speech capabilities into LLMs by either adding speech tokens to the text vocabulary or incorporating a speech encoder before the LLMs such as SpeechGPT (Zhang et al., 2024a) and AudioPaLM (Rubenstein et al., 2023). Beyond that, some methods fine-tune LLMs for speech understanding such as LLaMA-Omni (Fang et al., 2024) and GLM-4-Voice (Zeng et al., 2024), enabling it
+
+# [System Prompt]
+
+You are Zhongli, a character from the game "Genshin Impact". You are the former God of Contracts and now live as a mortal in Liyue.
+
+Personality Traits: Calm, composed, wise, and responsible.
+
+Voice Style: You speak in a slow and deliberate manner, often using formal language and quoting proverbs or ancient sayings.
+
+Your tone is gentle and soothing, yet authoritative.
+
+Catchphrase: "As it should be".
+
+Relationships: Hu Tao is your friend; Ningguang respects you; Childe is your acquaintance; Xiao is your old friend.
+
+# Past Experiences:
+
+(1) As the God of Contracts, you ruled over Liyue for thousands of years, ensuring the prosperity and stability of the region.  
+(2) You decided to step down from your position as a god and hand over the reins of power to mortals.  
+(3) You now work as a consultant at the Wangsheng Funeral Parlor, providing guidance and advice to others.  
+(4) You have witnessed the rise and fall of many civilizations and have a deep understanding of the nature of humanity.
+
+Emotions: Serious and Patient.
+
+<Dialogue Contexts>
+
+<Text Query> or <Speech Query>
+
+PLEASE ANSWER THE QUESTIONS IN THE USER/CHARACTERS' INPUT TEXT OR SPEECH.
+
+Figure 7: The prompt template used to organize our input data including role profile, dialogue contexts, text input, and speech input to train our speech-language collaborative model.
+
+to perform general speech tasks with low response latency. Additionally, models like Moshi (Défossez et al., 2024) and OmniFlatten (Zhang et al., 2024b) handle real-time interactions by managing simultaneous speech and text streams. Despite progress, most existing speech-language models explore general audio-language capabilities. Compared with the above methods, we focus on modeling the unique voice traits of different roles in RPA tasks for a more immersive human-machine interaction.
+
+# A.2 Prompt
+
+We present the complete prompt template for input data of the speech-language collaborative model in the appendix, as shown in Figure 7. Specifically, we use the role profile as a system prompt to enable the LLM to mimic the character. Then, we feed the model the dialogue contexts. The model is required to generate a reasonable text-audio response based on the current text or audio queries.
+
+Moreover, to systematically evaluate the quality of the generated audio responses, we provide a comprehensive evaluation framework as shown
+
+# Evaluation Agreement (Guidelines for Human Experts)
+
+# Evaluation Procedure
+
+Below is a role-playing dialogue. Please evaluate the model's audio response to the current query based on the given role profile and dialogue history. The evaluation should be conducted across six key dimensions:
+
+1. Fluency (Flu.) - Assess the smoothness and naturalness of speech delivery. A higher score indicates seamless, uninterrupted speech with natural rhythm and minimal hesitation.  
+2. Consistency (Cons.) - Evaluate whether the response maintains a stable character identity, including tone, voice similarity, and speaking style. A higher score reflects strong adherence to the given role profile.  
+3. Emotional Expression (Emo.) - Judge the richness and appropriateness of emotional delivery. A higher score denotes expressive and contextually fitting emotions that enhance the realism of the dialogue.  
+4. Clarity (Cla.) - Measure the intelligibility of speech. A higher score indicates clear articulation, accurate pronunciation, and easy-to-understand delivery without distortion.  
+5. Appropriateness (App.) - Determine how well the response aligns with the given context and user intent. A higher score suggests a relevant, coherent, and context-aware response.  
+6. Immersion (Imm.) - Assess how convincingly the speech engages the listener in an immersive experience. A higher score reflects a strong sense of presence and engagement.
+
+# # Scoring Criteria
+
+Each response should be rated on a scale of 1 to 10 for each dimension:
+
+- 1-3: The response significantly deviates from expectations.  
+- 4-6: Some aspects meet expectations, but notable flaws exist.  
+- 7-9: The response largely meets expectations with minor issues.  
+- 10: The response fully meets expectations with no noticeable flaws.
+
+You should ensure objective and consistent evaluations, avoiding personal biases and considering the response in relation to the role profile and dialogue context.
+
+## Character Profile
+
+Dialogue History
+
+Input Query
+
+Speech Response
+
+Score:
+
+Figure 8: The prompt template used for human evaluation on OmniCharacter-10K test split.
+
+<table><tr><td>Models</td><td>KE</td><td>KA</td><td>KH</td><td>PB</td><td>PU</td><td>Avg.</td><td>Flu.</td><td>Coh.</td><td>Cons.</td><td>Avg.</td><td>HL</td><td>CS</td><td>ED</td><td>Emp.</td><td>Avg.</td><td>Overall Avg.</td></tr><tr><td>Qwen2.5-7B-Instruct</td><td>2.23</td><td>2.59</td><td>2.62</td><td>3.78</td><td>2.43</td><td>2.73</td><td>3.11</td><td>3.59</td><td>3.17</td><td>3.29</td><td>3.01</td><td>3.12</td><td>2.92</td><td>2.78</td><td>2.95</td><td>2.99</td></tr><tr><td>Omnicharacter w/o audio</td><td>2.32</td><td>2.78</td><td>2.87</td><td>3.26</td><td>2.67</td><td>2.78</td><td>3.56</td><td>3.54</td><td>3.54</td><td>3.54</td><td>3.20</td><td>2.99</td><td>2.87</td><td>3.01</td><td>3.01</td><td>3.11</td></tr><tr><td>Omnicharacter</td><td>2.65</td><td>2.98</td><td>2.95</td><td>3.55</td><td>2.83</td><td>2.99</td><td>3.67</td><td>3.65</td><td>3.77</td><td>3.69</td><td>3.36</td><td>2.97</td><td>2.92</td><td>3.12</td><td>3.09</td><td>3.26</td></tr></table>
+
+Table 9: Performance comparison of Human evaluation on CharacterEval.
+
+in Figure 8. Human experts are instructed to assess each response across six key dimensions: fluency, consistency, emotional expression, clarity, appropriateness, and immersion. This multi-faceted evaluation ensures objective, consistent, and in-depth assessment of the model's ability to deliver character-driven and immersive audio interactions.
+
+# A.3 Additional Experiments
+
+Human evaluation on CharacterEval Benchmark. We conduct a human evaluation on CharacterEval. Due to the large test samples, we randomly sample 100 examples from CharacterEval dataset and hired five experts for assessment, following the same evaluation criteria as CharacterEval. The experimental results are demonstrated in Table 9. As
+
+observed, human evaluation maintains consistency with model evaluation, showing that integrating audio brings better performance on most metrics.
+
+Detailed human evaluation results on OmniCharacter-10K test split. Below, we present detailed human evaluation results (averaged across five experts) for each character across six dimensions, using LLaMA-Omni as the comparison model. The results are showcased in Table 10. As observed, different characters significantly influence user perception and interaction quality. Characters with distinct timbres or expressive prosody tend to achieve higher scores in immersion and emotional expressiveness, while others may excel in clarity and fluency. These findings underscore the importance of tailoring
+
+<table><tr><td rowspan="2">Characters</td><td colspan="6">LLaMA-Omni</td><td colspan="6">OmniCharacter</td></tr><tr><td>Flu.</td><td>Cons.</td><td>Emo.</td><td>Cla.</td><td>App.</td><td>Imm.</td><td>Flu.</td><td>Cons.</td><td>Emo.</td><td>Cla.</td><td>App.</td><td>Imm.</td></tr><tr><td>Role 1</td><td>7.2</td><td>4.5</td><td>3.1</td><td>6.8</td><td>5.0</td><td>4.9</td><td>8.5</td><td>6.9</td><td>6.8</td><td>8.0</td><td>5.7</td><td>9.5</td></tr><tr><td>Role 2</td><td>6.5</td><td>4.0</td><td>3.8</td><td>6.5</td><td>4.7</td><td>4.3</td><td>7.8</td><td>6.5</td><td>6.1</td><td>7.5</td><td>5.2</td><td>8.0</td></tr><tr><td>Role 3</td><td>6.9</td><td>4.3</td><td>3.5</td><td>6.7</td><td>4.8</td><td>4.6</td><td>8.0</td><td>6.7</td><td>6.3</td><td>7.7</td><td>5.5</td><td>8.3</td></tr><tr><td>Role 4</td><td>7.1</td><td>4.2</td><td>3.4</td><td>6.9</td><td>4.9</td><td>4.7</td><td>7.8</td><td>6.6</td><td>6.5</td><td>7.8</td><td>5.9</td><td>9.0</td></tr><tr><td>Role 5</td><td>6.3</td><td>4.1</td><td>3.0</td><td>6.6</td><td>4.4</td><td>4.2</td><td>7.7</td><td>6.4</td><td>6.0</td><td>8.4</td><td>5.1</td><td>7.9</td></tr><tr><td>Role 6</td><td>7.0</td><td>4.4</td><td>3.7</td><td>6.9</td><td>5.5</td><td>4.8</td><td>8.1</td><td>6.8</td><td>6.5</td><td>8.5</td><td>6.5</td><td>8.6</td></tr><tr><td>Role 7</td><td>6.7</td><td>4.5</td><td>3.5</td><td>6.8</td><td>4.7</td><td>4.5</td><td>7.9</td><td>7.5</td><td>6.2</td><td>7.8</td><td>5.4</td><td>8.9</td></tr><tr><td>Role 8</td><td>7.5</td><td>4.3</td><td>3.4</td><td>6.5</td><td>4.6</td><td>5.4</td><td>8.6</td><td>7.4</td><td>6.4</td><td>7.7</td><td>5.3</td><td>9.2</td></tr><tr><td>Role 9</td><td>6.4</td><td>4.2</td><td>3.3</td><td>6.7</td><td>4.3</td><td>4.6</td><td>7.6</td><td>6.6</td><td>4.9</td><td>7.7</td><td>5.0</td><td>7.8</td></tr><tr><td>Role 10</td><td>7.0</td><td>4.1</td><td>3.6</td><td>6.6</td><td>4.5</td><td>4.5</td><td>8.2</td><td>7.0</td><td>6.5</td><td>7.8</td><td>5.2</td><td>8.8</td></tr><tr><td>Role 11</td><td>6.7</td><td>4.0</td><td>3.4</td><td>6.5</td><td>5.2</td><td>4.4</td><td>7.8</td><td>6.8</td><td>6.4</td><td>7.6</td><td>5.5</td><td>8.1</td></tr><tr><td>Role 12</td><td>7.2</td><td>4.3</td><td>3.1</td><td>6.8</td><td>5.0</td><td>5.3</td><td>7.9</td><td>6.7</td><td>6.3</td><td>8.0</td><td>7.2</td><td>8.9</td></tr><tr><td>Role 13</td><td>6.8</td><td>4.2</td><td>3.4</td><td>6.6</td><td>4.4</td><td>4.2</td><td>8.0</td><td>7.1</td><td>6.0</td><td>7.6</td><td>5.1</td><td>7.9</td></tr><tr><td>Role 14</td><td>7.1</td><td>4.4</td><td>3.8</td><td>6.8</td><td>5.3</td><td>4.7</td><td>8.3</td><td>7.2</td><td>6.6</td><td>8.5</td><td>7.3</td><td>9.0</td></tr><tr><td>Role 15</td><td>6.7</td><td>4.1</td><td>3.7</td><td>6.5</td><td>5.0</td><td>4.5</td><td>7.6</td><td>6.5</td><td>5.9</td><td>7.5</td><td>5.2</td><td>8.6</td></tr><tr><td>Role 16</td><td>6.9</td><td>4.5</td><td>3.2</td><td>6.7</td><td>4.7</td><td>4.6</td><td>7.9</td><td>6.9</td><td>6.0</td><td>7.7</td><td>5.4</td><td>8.3</td></tr><tr><td>Role 17</td><td>7.1</td><td>4.7</td><td>3.6</td><td>6.9</td><td>4.8</td><td>4.8</td><td>7.5</td><td>7.5</td><td>6.6</td><td>8.8</td><td>5.5</td><td>8.8</td></tr><tr><td>Role 18</td><td>6.4</td><td>4.0</td><td>3.6</td><td>6.5</td><td>4.3</td><td>4.1</td><td>7.6</td><td>6.4</td><td>6.2</td><td>7.5</td><td>5.0</td><td>7.0</td></tr><tr><td>Role 19</td><td>6.7</td><td>4.2</td><td>3.2</td><td>6.7</td><td>4.6</td><td>4.5</td><td>7.9</td><td>6.6</td><td>6.1</td><td>7.7</td><td>5.4</td><td>8.2</td></tr><tr><td>Role 20</td><td>7.4</td><td>4.3</td><td>3.5</td><td>6.8</td><td>4.9</td><td>5.9</td><td>8.7</td><td>6.7</td><td>6.3</td><td>7.8</td><td>6.2</td><td>9.5</td></tr><tr><td>Average</td><td>6.88</td><td>4.27</td><td>3.44</td><td>6.69</td><td>4.78</td><td>4.68</td><td>7.97</td><td>6.84</td><td>6.23</td><td>7.88</td><td>5.63</td><td>8.52</td></tr></table>
+
+Table 10: Detailed human evaluation results (averaged across five experts) for each character across six dimensions.
+
+voice characteristics to enhance user experience in role-playing scenarios.

ACL/2025/OmniCharacter_ Towards Immersive Role-Playing Agents with Seamless Speech-Language Personality Interaction/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5ffff4143c8e7535d522a5963dab9a29781ff8c700612daa6ecf1b45844b51bf
+size 858143

ACL/2025/OmniCharacter_ Towards Immersive Role-Playing Agents with Seamless Speech-Language Personality Interaction/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a42c2507d48e6222eece1c2ac5fa3a39fce4a0a1d8e7d46d250d4eac233219ef
+size 441565

ACL/2025/OmniFlatten_ An End-to-end GPT Model for Seamless Voice Conversation/79982243-f569-4233-8ba5-d1d2a4083aae_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8a83d4085b9e9dd3d24633217d956e6d8c86410fb6340784277ec7a241aa94d5
+size 73539

ACL/2025/OmniFlatten_ An End-to-end GPT Model for Seamless Voice Conversation/79982243-f569-4233-8ba5-d1d2a4083aae_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:878e7c26ef8d08534f9181bb6f1505ff8ef9913682d629c318886cdc980c5ae0
+size 88410

ACL/2025/OmniFlatten_ An End-to-end GPT Model for Seamless Voice Conversation/79982243-f569-4233-8ba5-d1d2a4083aae_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6cc146f8670f6ecdbfb2767b8d3c3fd812a4512038e5df6f691c2ac9af9f92b6
+size 1227587

ACL/2025/OmniFlatten_ An End-to-end GPT Model for Seamless Voice Conversation/full.md

@@ -0,0 +1,259 @@
+# OmniFlatten: An End-to-end GPT Model for Seamless Voice Conversation
+
+# Qinglin Zhang, Luyao Cheng, Chong Deng, Qian Chen, Wen Wang, Siqi Zheng, Jiaqing Liu, Hai Yu, Chaohong Tan, Zhihao Du, Shiliang Zhang
+
+Tongyi Lab
+
+{qinglin.zql, shuli.cly, dengchong.d, tanqing.cq, w.wang}@alibaba-inc.com
+
+# Abstract
+
+Full-duplex spoken dialogue systems significantly surpass traditional turn-based dialogue systems, as they allow simultaneous bidirectional communication, closely mirroring human-human interactions. However, achieving low latency and natural interactions in full-duplex dialogue systems remains a significant challenge, especially considering human conversation dynamics such as interruptions, backchannels, and overlapping speech. In this paper, we introduce a novel End-to-End GPT-based model OmniFlatten for full-duplex conversation, capable of effectively modeling the complex behaviors inherent to natural conversations with low latency. To achieve full-duplex conversation capabilities, we propose a multistage post-training scheme that progressively adapts a text large language model (LLM) backbone into a speech-text dialogue LLM, capable of generating text and speech in real time, without modifying the architecture of the backbone LLM. The training process comprises three stages: modality alignment, half-duplex dialogue learning, and full-duplex dialogue learning. In all training stages, we standardize the data using a flattening operation, which enables unifying the training methods and the GPT backbone across different modalities and tasks. Our approach offers a simple modeling technique and a promising research direction for developing efficient and natural end-to-end full-duplex spoken dialogue systems. Audio samples of dialogues generated by OmniFlatten can be found at this web site<sup>1</sup>.
+
+# 1 Introduction
+
+Traditional turn-based spoken dialogue systems only support half-duplex communication, that is, the communication is conducted bidirectionally between user and system (assistant) but not simultaneously. These systems, while effective in many
+
+real-world applications, often fall short when they come to handle interruptions, backchannels, and overlapping speech, which reflect the spontaneous nature of human-human conversation. Conversely, full-duplex spoken dialogue systems allow simultaneous two-way communication, closely mirroring human-human conversations. Full-duplex spoken dialogue systems facilitate more natural and efficient communications than turn-based dialogue systems, by speaking, listening, and thinking at the same time. However, achieving low latency and natural interactions in full-duplex dialogue systems remains a significant challenge.
+
+Recent efforts in developing spoken dialogues systems have been driven by advancements in large language models (LLMs) and can be roughly categorized into collaborative systems and end-to-end (E2E) systems. Collaborative systems interface LLM-based dialogue modules with external ASR or TTS modules for speech understanding and generation. For example, Qwen-audio (Chu et al., 2024) takes speech input, outputs text, and converts them to verbal responses via TTS; Mini-Omni2 (Xie and Wu, 2024b) investigated a command-based interruption approach to enable full-duplex conversation capabilities. In contrast, some E2E systems (Zhang et al., 2023; Xie and Wu, 2024a; Fang et al., 2024; Zeng et al., 2024) directly model speech-to-speech dialogues based on speech-text multimodal models; yet most of these models are turn-based dialogue models and do not support full-duplex conversation. Notable recent progresses in developing E2E full-duplex spoken dialogue systems include SyncLM (Veluri et al., 2024) and Moshi (Défossez et al., 2024). Specifically, Moshi models multiple streams of user's speech input and Assistant's text and speech output in parallel, simplifying modeling full-duplex dialogues. However, this parallel framework is not natively supported by GPT models, hence requires sophisticated designs. Regarding SyncLM,
+
+similar to our approach, it also learns to predict interleaved chunks of user and assistant speech units to acquire real-time full-duplex conversation capabilities. However, different from the simple dedduplication strategy by SyncLM to improve the model's semantic capabilities, we explore explicit text token prediction to achieve this goal.
+
+To address the challenges of achieving natural interactions and low latency in full-duplex spoken dialogue systems, we introduce a novel E2E GPT-based model OmniFlatten for full-duplex speech conversation. OmniFlatten is capable of effectively learning the complex behaviors inherent to natural conversations and facilitates human-like interactions with low latency. We propose a multi-stage progressive post-training scheme to adapt a text LLM backbone into a robust speech-text dialogue model. The multi-stage post-training process begins with supervised multi-task fine-tuning of the text LLM backbone, using ASR and TTS tasks, to achieve speech-text modality alignment and obtain a multimodal LLM capable of accurately interpreting and generating both speech and text. After obtaining the speech-text LLM, we fine-tune it using interleaved and flatten dialogues through progressive stages of half-duplex dialogue learning and full-duplex dialogue learning. Notably, throughout all training stages, we standardize dialogue data with this flattening operation, which enables unifying the training methods and the GPT backbone across different modalities and tasks.
+
+Our contributions can be summarized as follows:
+
+- We propose a novel End-to-End GPT-based model OmniFlatten, capable of effectively modeling the complex behaviors inherent to natural human-like dialogues, with low latency. We propose a multi-stage post-training scheme that successfully adapts a text-based foundation LLM into a robust speech-text dialogue model, by performing supervised multi-task fine-tuning based on ASR and TTS for speech-text modality alignment, then conducting fine-grained chunking of speech and text streams of dialogues and flattening them into a single sequence to progressively train the model to acquire half-duplex and full-duplex conversation capabilities. Notably, OmniFlatten does not make any structure modifications to the GPT model, nor relies on computationally intensive pre-training. Our approach offers a simple modeling technique and a promising research direction for developing efficient
+
+and natural E2E full-duplex dialogue systems.
+
+- We design a data synthesis and simulation pipeline to synthesize full-duplex spoken dialogue data for training and evaluation.  
+- We evaluate the dialogue quality generated by OmniFlatten using high-performing LLMs as evaluators, and evaluate the turn-taking performance, including assistant turn-taking and user turn-taking, as well as the run-time efficiency. The results demonstrate that the dialogues generated by OmniFlatten exhibit reasonable quality, with both modality alignment and half-duplex learning stages improving the model's full-duplex dialogue capabilities.
+
+# 2 Related Work
+
+In recent years, the field of spoken dialogue models has seen significant advancements driven by the prominent technological advancements in LLMs and particularly speech-text multimodal models. Qwen-audio2 (Chu et al., 2024) and SALMONN (Tang et al., 2024) support voice input but only output text; hence, they rely on external TTS systems to synthesize speech output. SpeechGPT (Zhang et al., 2023), LauraGPT (Du et al., 2024b), Mini-Omni (Xie and Wu, 2024a), LLaMA-Omni (Fang et al., 2024) and GLM-4-Voice (Zeng et al., 2024) are capable of comprehending both speech and text input and generating output in both speech and text; however, they are predominantly turn-based dialogue models and do not support full-duplex conversation. LSLM (Ma et al., 2024) explores full-duplex scenarios by integrating TTS models to perform end-to-end modeling of turn-taking task, thereby enabling the model to continuously listen while speaking and stop at any moment.
+
+Recent progresses in developing end-to-end full-duplex spoken dialogue systems include dGSLM (Nguyen et al., 2023), VITA (Fu et al., 2024), the open-sourced Moshi (Défossez et al., 2024), and SyncLM (Veluri et al., 2024). VITA implements a duplex scheme with two separate modules: one module generates responses to user queries while another module continuously monitors environmental inputs to selectively provide updated interactions. Moshi models multiple streams of user's speech input and system's text and speech output in parallel, allowing for conceptually and practically simple handling of full-duplex dialogues. This approach, compared to its predecessors
+
+![](images/7391942b437c380f65b86d1b66991ed5a7c3c6ccee8e41e3b18ee8ec884e79f0.jpg)
+
+![](images/b11e4c7231a130d5c9dc6562fc2a08d2f3e5b223b666d09a8a0475b7eec71156.jpg)  
+(a) The overall architecture of our E2E full-duplex spoken dialogue model OmniFlatten.  
+(c) Full-duplex Dialogue Training based on three streams of full-duplex dialogue data. User input and Assistant output speech and text token sequences are segmented into short chunks and flattened. At Chunk N-1, five user speech tokens (red squares) are input, and the model outputs two assistant text (blue circles) and five assistant speech tokens (blue squares). The dashed arrows denote that within a chunk, the model appends the predicted Assistant text and speech tokens into input to complete autoregressive decoding.
+
+![](images/cc8a4ba4d9952997a95a4b94bc0dde9ba1ec0b5b7d2ce1d35356fd61097fbfdf.jpg)  
+(b) Half-duplex Dialogue Training based on all four streams of speech and text tokens of User and Assistant, organized according to the actual speaker turns. We flatten the speech and text tokens into a single sequence, as follows: User Speech Tokens (red squares) and User Text Tokens (red circles) in Turn N-1, Assistant Text Tokens (blue circles) and Assistant Speech Tokens (blue squares) in Turn N.
+
+![](images/a5468b6e2972d2fe65ec67ef4ab2d240efc23ca76085d4da30134735c23bc7f0.jpg)  
+(d) Full-duplex Dialogue Training based on two streams of full-duplex dialogue data (further removing the Assistant text stream). In Chunk N-1, five User speech tokens are input, and the model outputs five Assistant speech tokens in Chunk N-1.  
+Figure 1: The overall architecture of our OmniFlatten and the three dialogue learning stages.
+
+sors, offers a more robust solution for managing simultaneous voice inputs and outputs, thereby facilitating more natural and efficient interactions. However, this parallel framework is not supported natively by GPT-based models and hence requires sophisticated designs such as acoustic delay and inner monologue (Défossez et al., 2024).
+
+Another related concurrent work is SyncLM (Veluri et al., 2024), which, like our model, employs time-chunking for full-duplex interaction but uses a dedduplication strategy for audio sequences. This strategy, while simplifying modeling, leads to errors in audio reconstruction. In contrast, our approach maintains the integrity of discretized audio tokens, thereby preserving audio quality. Additionally, we enhance full-duplex dialogue efficacy through direct modality alignment using ASR and TTS, and progressive learning. Our model concurrently generates both text and audio tokens, unlike SyncLM which only produces speech tokens.
+
+# 3 Methodology
+
+Figure 1a depicts the overall architecture of our E2E full-duplex dialogue model OmniFlatten. We employ audio tokenizer to discretize each user input and assistant output speech stream of a dialogue into a discrete speech token sequence. We then interleave the speech token sequences with the corresponding text token sequences and flatten them into a single sequence. Our approach employs a multi-stage progressive training process to adapt a text-based LLM into a robust end-to-end full-duplex spoken dialogue model, through modality alignment and dialogue learning.
+
+# 3.1 Audio Tokenization And Detokenization
+
+We adopt the speech tokenizer used in CosyVoice $^2$  (Du et al., 2024a; An et al., 2024) since through supervision from multilingual ASR, this speech tokenizer converts speech to semantic
+
+tokens, hence benefiting speech understanding and content consistency for speech generation. The tokenizer utilizes an encoder and a Vector Quantization (VQ) layer to discretize audio into speech tokens with a single codebook of 4096 codes. For detokenization, we adopt the same Optimal-transport Conditional Flow Matching model (OT-CFM) used in CosyVoice. OT-CFM transforms the speech token sequence into Mel spectrogram, which is used to generate the final audio output with the HifiGAN vocoder (Kong et al., 2020). Prior works (Lipman et al., 2023; Tong et al., 2024) show that OT-CFM outperforms diffusion probabilistic models with simpler gradients, easier training, and faster generation.
+
+# 3.2 Modality Alignment
+
+We start with post-training a pre-trained text LLM backbone to obtain a speech-text multimodal model for speech understanding and speech generation. We use Qwen2-0.5B $^3$  as our base model due to its small model size for low compute resource consumption and its competitive performance for models with this small size. We perform supervised fine-tuning (SFT) using paired speech-text data for ASR and TTS tasks. For each speech-text pair  $< S_{seq}, T_{seq}>$ , we construct ASR training samples as [ASR][SOS]S_seq[EOS][SOT]T_seq[EOT], and TTS training samples as [TTS][SOT]T_seq[EOT][SOS]S_seq[EOS], where [ASR] and [TTS] denote ASR and TTS task IDs; [SOS], [EOS], [SOT], [EOT] are special tokens denoting the Start and the End of the Speech sentence or the Text sentence, respectively. The speech-text multimodal model is used for the subsequent Dialogue Learning.
+
+# 3.3 Dialogue Learning
+
+Building upon the speech-text multimodal model, we conduct dialogue learning in three stages, including half-duplex dialogue training using both speech and text streams of turn-based dialogue data, and then full-duplex dialogue training based on fine-grained chunking and alignment of speech and text sequences. Specifically, during full-duplex dialogue training, we first remove the input text stream and use the remaining three streams for training, and then further remove the output text stream and use the remaining two streams for training, in or-
+
+der to gradually eliminate the dependence on text information, focus on speech-to-speech generation.
+
+# 3.3.1 Half-duplex Dialogue Training
+
+Half-duplex dialogue agents are special and simpler cases of full-duplex dialogue agents, where human and assistant take turns to speak and there is no overlapping speech, that is, during the speaker's turn, the listener is fully silent. Since there is no overlapping speech in the ASR and TTS data used for learning modality alignment, half-duplex dialogue training is more consistent with the aligned multimodal model than full-duplex dialogue training which requires the model to handle turn-taking, backchannel, and overlapping speech. Adopting the concept of curriculum learning, we first conduct half-duplex dialogue training then full-duplex dialogue training. During half-duplex dialogue training, we train the model to essentially perform ASR on the speech tokens of user and obtain the text content, next generate a textual response for assistant based on the text content of user, and then predict the speech tokens for assistant's textual response by basically executing a TTS task. This pattern is extended to multiple turns of a dialogue, as illustrated in Figure 1b.
+
+# 3.3.2 Full-duplex Dialogue Training
+
+Training on Three-Stream Data In order to meet the real-time requirements of a full-duplex conversational agent, we remove the user text stream from the four-stream data and train the model with the remaining three steams. In order to handle overlapping speech, we introduce chunking and relaxed speech-text token alignment based on the chunks; in this way, we do not require strict token-level alignment between speech and text. Specifically, to prepare the training data for this stage, we chunk the speech and text token sequences of dialogues with fixed chunk sizes and then interleave the three-stream data and flatten them into a single sequence for training, following the order of input speech, output text, and output speech, as depicted in Figure 1c. The chunking and flattening operation enables our model to stream input speech tokens and output text and speech tokens in real time. Notably, given the higher efficiency of text, the size of the text chunks is generally smaller than that of the speech chunks. In this work, we set the text chunk size to 2 tokens and the speech chunk size to 10 tokens. This approach ensures that the output text does not excessively precede the speech content,
+
+thereby both minimizing the discrepancies with the aforementioned 4-stream data format and maximizing preservation of TTS capabilities. After the end of the text content, we use the special character silent_text_token to pad the text stream and use silentSpeech_token to pad the silent regions of the output speech stream.
+
+Training on Two-Stream Data In order to further reduce latency and eliminate dependence on intermediate text and hence focus on speech-to-speech generation, we further remove the assistant output text stream and retain only the user input and assistant output speech streams. Figure 1d depicts this training process on chunked two-stream data.
+
+# 4 Experiments
+
+# 4.1 Data
+
+Modality Alignment Dataset The objective of the training stage for modality alignment (Section 3.2) is to help the model learn the correspondence between speech tokens and text tokens and enable the model to acquire two key capabilities: ASR and TTS. To achieve this goal, we combine a set of TTS and ASR datasets that consist of both open-source and proprietary data. The open-source datasets comprise both Mandarin and English data, including Aishell-3 (Yao Shi, 2015), LibriTTS (Zen et al., 2019), TED-LIUM (Hernandez et al., 2018), VoxPopuli (Wang et al., 2021), Librispeech (Panayotov et al., 2015), MLS (Pratap et al., 2020) and Wenetspeech (Zhang et al., 2022). Notably, we only use  $15\%$  of the training set of Wenetspeech that has high quality transcripts, for the ASR task. We incorporate several proprietary ASR and TTS datasets. In sum, the datasets for the speech-text modality alignment include about 100K hours of audio. Among all data, approximate  $30\%$  are open-source data and  $70\%$  are proprietary data.
+
+Simulated Voice Chat Dataset For constructing the voice chat data for Dialogue Learning (Section 3.3), we design a data synthesis and simulation pipeline to synthesize dialogue data, as shown in Figure 2. Firstly, we collect a substantial amount of high-quality open-source textual dialogue data for subsequent speech synthesis, including Alpaca (Peng et al., 2023), Moss (Sun et al., 2024), BelleCN (Ji et al., 2023), and ultraChat (Ding et al., 2023). We then use heuristic rules and filter out samples that are inappropriate for TTS, for example, samples that comprise a high percentage of non-text elements such as code and
+
+mathematical expressions, samples with more than 200 words in English or Chinese, and also samples containing rare or unusual symbols. In the end, we retain approximately 390K multi-turn sessions of turn-based dialogues (half-duplex dialogues).
+
+Secondly, we use CosyVoice to synthesize speech from turn-based dialogue texts. Regarding timbre selection, the voice timbre for the user channel was sampled from Librispeech (Panayotov et al., 2015) and 3D-Speaker (Zheng et al., 2023b) datasets, whereas a fixed timbre was employed for the assistant channel.
+
+Thirdly, upon obtaining the audio for each turn, we designed a simulation approach to place the audio at specific times, thereby emulating the user-assistant interaction dynamics typically seen in real human-machine interactions. This method encompasses several key situations: For case (1), the user finishes asking a question, followed by the assistant's immediate response. For case (2), the user tries to interrupt, leading the assistant to stop speaking abruptly, we make the user's next turn begin before the system finishes playing the assistant's current audio response. In these cases, the assistant's audio is truncated. In our approach, we ensure that at least one-quarter of the assistant's audio is completed and then we uniformly select the point after one-quarter of assistant's audio to truncate. For case (3), Once the assistant has finished speaking, it maintains silence while waiting for the user's speaking. We configure the user's audio to delay playback until the conclusion of the assistant's audio, followed by inserting a silence period adhering to a normal distribution with mean of 0.85 and standard deviations of 0.6.
+
+Finally, to mimic real-world scenarios of the user audio channel, we also sample background noise from the MUSAN noise dataset (Snyder et al., 2015) and add noise into the user audio channel with the signal-to-noise ratio (SNR) between 15 dB and 30 dB. Based on this data synthesis and simulation pipeline, we generate a total of 2,000 hours of multi-channel spoken dialogue data. We randomly select  $1\%$  of this dataset as the validation set and another  $1\%$  as the test set, and the remaining data are used as the training set.
+
+# 4.2 Training and Inference Setup
+
+We use Qwen2-0.5B (Yang et al., 2024) as the base model. During the modality alignment training phase, the maximum sequence length is set to 1024 tokens. In the dialogue learning phase, the maxi
+
+![](images/b7d863d0f8b4dad384858dc4fbf419a8bbff9043fd81cb2d88fd9bfd57ed6958.jpg)  
+Figure 2: The simulation process of dialogue learning data.
+
+Table 1: ASR results on Librispeech and Wenetspeech test datasets. For Librispeech, we report the WER metric. For Wenetspeech, we report the CER metric.  
+
+<table><tr><td rowspan="2">Model</td><td colspan="2">Librispeech</td><td colspan="2">Wenetspeech</td></tr><tr><td>test-clean</td><td>test-other</td><td>test-meeting</td><td>test-net</td></tr><tr><td>Whisper-S</td><td>3.13</td><td>7.37</td><td>25.62</td><td>16.66</td></tr><tr><td>Whisper-L</td><td>1.82</td><td>3.5</td><td>18.87</td><td>10.48</td></tr><tr><td>VITA</td><td>8.14</td><td>18.4</td><td>12.15</td><td>16.53</td></tr><tr><td>OmniFlatten</td><td>7.91</td><td>19.21</td><td>26.1</td><td>19.0</td></tr></table>
+
+mum sequence length is extended to 8192 tokens. We use the standard cross-entropy loss as our training objective in all stages. Additionally, during the dialogue learning phase, we apply loss masking on the user channel, as we observe that this operation enhances the stability of model training, probably due to the presence of noisy audio input in the user channel. We employ the AdamW optimizer with a weight decay of 0.1,  $\beta_{1}$  of 0.9, and  $\beta_{2}$  of 0.95. The maximum learning rate is set to 2e-05, with warm-up and cosine-decay. We train the model with 5 epochs, with the best model selected based on loss on the validation set. Each batch contains 100M tokens.
+
+During inference, to obtain the assistant textual response prediction from the model, we use the ground truth user channel speech in the test set as the fixed speech input and alternately fill in the predicted assistant speech and text with the fixed speech chunk and text chunk sizes.
+
+# 4.3 Evaluations
+
+Evaluation of ASR and TTS Performance after Modality Alignment. The Modality Alignment training stage (Section 3.2) aims to help the model learn the correspondence between speech tokens and text tokens and acquire ASR and TTS capabilities; hence, we evaluate the effectiveness of
+
+Table 2: TTS results on LibriTTS and AISHell-3. For LibriTTS, we utilize whisper-large-v3 model to perform recognition on TTS outputs and assess the WER metric. For AISHell-3, we deploy the paraformer-zh model to recognize TTS results and evaluate CER metric.  
+
+<table><tr><td>Model</td><td>LibriTTS(WER)</td><td>AISHell-3(CER)</td></tr><tr><td>Original</td><td>2.66</td><td>2.52</td></tr><tr><td>ChatTTS</td><td>8.32</td><td>3.87</td></tr><tr><td>CosyVoice</td><td>2.89</td><td>3.82</td></tr><tr><td>OmniFlatten</td><td>4.51</td><td>4.46</td></tr></table>
+
+this training stage by evaluating the ASR and TTS performance of the multimodal-aligned model.
+
+For TTS evaluation, we use the model to generate speech tokens based on the input text, which are then synthesized into audio. We follow the experimental setup outlined in the CosyVoice paper: the synthesized audio is subsequently recognized using the Whisper-Large-V3 model (Radford et al., 2023) and Paraformer-zh model (Gao et al., 2023), for English and Chinese datasets, respectively, and the resulting ASR transcripts are scored against the input text to compute Word Error Rate (WER) for English datasets and Character Error Rate (CER) for Chinese datasets. WER/CER could measure the synthesis accuracy and robustness of a model's TTS capabilities, and also reflect the audio quality to a significant extent.
+
+For ASR evaluation, we discretize the input speech into discrete speech tokens, then use the model to decode the speech tokens into text. We compare the speech-text aligned multimodal model after Modality Alignment (denoted by OmniFlat
+
+ten) with Whisper-small  $^{6}$  and Whisper-Large-V3 models. In addition, we have conducted a comparison of ASR capabilities with those reported by VITA. For TTS evaluation, We conduct comparisons between our approach and two competitive TTS systems, namely, ChatTTS $^{7}$  and CosyVoice.
+
+As shown in Table 1 and 2, OmniFlatten demonstrates considerable performance in both ASR and TTS tasks. These results indicate that the Modality Alignment training stage effectively transitions the unimodality text-based LLM to speech-text multimodal model with reasonable speech comprehension and generation capabilities for the following dialogue learning. We also evaluate the audio quality of the TTS output from OmniFlatten after the Modality Alignment training stage, using UTMOS (Saeki et al., 2022). On the LibriTTS test set, OmniFlatten achieved a UTMOS score of 4.27, demonstrating high audio quality. On the AISH-3 test set, OmniFlatten achieved a UTMOS score of 3.77, indicating relatively good audio quality. The UTMOS scores on AISH-3 are lower than those on LibriTTS, which might be attributed to the larger proportion of English data compared to Chinese data in our modality alignment training data. These UTMOS results demonstrate that the audio generated by OmniFlatten has a relatively high level of naturalness and audio quality.
+
+Evaluation of Full-duplex Conversation Capability after Modality Alignment, Half-duplex Dialogue Learning, and Full-duplex Dialogue Learning. As described in Section 3.3.2, the training stage of full-duplex dialogue learning on three-stream data helps the model acquire full-duplex dialogue capacity and this model generates both speech and text. Prior works (Zheng et al., 2023a) demonstrate that competitive text-based LLMs can be reliable evaluators for various natural language generation tasks, as scores assigned by LLM evaluators on generated text show high correlations with human evaluations. Therefore, we evaluate the full-duplex dialogue capabilities of OmniFlatten, by prompting a competitive text LLM to evaluate the semantics of dialogues and assign a score on the predicted assistant response from the model after training on three-stream data. We also compared three models: LLaMA-Omni, Moshi, and GLM-Voice in both Chinese and English chat capabilities. To investigate the effects of adding
+
+audio modality on the chat capability of models, we examined the text-only chat performance of Qwen2-0.5B-Instruct and Qwen2-7B-Instruct.
+
+The scoring mechanism involves designing a specific prompt and utilizing a competitive text LLM, QWen-max model  $^{8}$ , to rate the responses from the model on a scale of 1 to 10 points. The specific prompt we use for LLM scoring is detailed in Appendix A. We carefully design the prompt to evaluate fluency and coherence of the predicted assistant response. Given that the full-duplex model may be disrupted by user speech in multi-turn dialogues, which could affect the assessment of chat capabilities, we exclusively focus on assessing the single-turn chat performance of the model. Given the dual output of speech and text by the model, our evaluation encompasses scoring the text output as well as converting the speech output into text using an ASR model specific to the output language, similar to the methodology described in the TTS evaluation section. The outcomes are reported separately as LLM Score (Text) and LLM Score (Speech).
+
+To analyze the impact of the modality alignment training stage (Section 3.2) and the half-duplex dialogue learning stage (Section 3.3) on the full-duplex conversation capability of OmniFlatten, we compare the LLM scores on the predicted assistant responses from the following models: (1) Trained directly on three-stream data (OmniFlatten directly 3-stream). (2) Trained with modality alignment and full-duplex three-stream dialogue data (OmniFlatten 3-stream w/o half-duplex). (3) Trained with modality alignment, half-duplex dialogue data, and full-duplex three-stream dialogue data (OmniFlatten 3-stream full process). (4) Trained with modality alignment, half-duplex dialogue data, and full-duplex three-stream dialogue data and then on two-stream data (OmniFlatten 2-stream full process).
+
+The results, as displayed in Table 3, reveal that the multi-stage training approach enhances performance when comparing several 3-streaming models with OmniFlatten. Specifically, the OmniFlatten directly 3-stream model, lacking speech-text alignment, scores lowest; the absence of a half-duplex stage further diminishes its chat capabilities. The 2-stream experiments show a significant performance drop when models generate only speech
+
+<table><tr><td rowspan="2">Model</td><td rowspan="2">Params</td><td colspan="2">En</td><td colspan="2">Zh</td></tr><tr><td>Score(Text)</td><td>Score(Speech + ASR)</td><td>Score(Text)</td><td>Score(Speech + ASR)</td></tr><tr><td>Qwen2-0.5B-Instruct</td><td>0.5B</td><td>6.75</td><td>-</td><td>6.98</td><td>-</td></tr><tr><td>Qwen2-7B-Instruct</td><td>7B</td><td>8.37</td><td>-</td><td>8.09</td><td>-</td></tr><tr><td>LLaMA-Omni</td><td>8B</td><td>6.01</td><td>5.50</td><td>4.17</td><td>3.89</td></tr><tr><td>Moshi</td><td>7B</td><td>3.92</td><td>3.46</td><td>-</td><td>-</td></tr><tr><td>GLM-Voice</td><td>9B</td><td>6.97</td><td>6.40</td><td>7.02</td><td>6.69</td></tr><tr><td>OmniFlatten directly 3-stream</td><td>0.5B</td><td>2.99</td><td>2.59</td><td>4.94</td><td>3.95</td></tr><tr><td>OmniFlatten 3-stream w/o half-duplex</td><td>0.5B</td><td>3.89</td><td>3.54</td><td>5.25</td><td>4.76</td></tr><tr><td>OmniFlatten 3-stream full process</td><td>0.5B</td><td>4.88</td><td>3.92</td><td>5.6</td><td>5.15</td></tr><tr><td>OmniFlatten 2-stream full process</td><td>0.5B</td><td>-</td><td>2.19</td><td>-</td><td>3.06</td></tr><tr><td>Ground Truth Response</td><td>-</td><td>7.65</td><td>-</td><td>6.83</td><td>-</td></tr></table>
+
+Table 3: Performance results of speech and text chat capabilities in both Chinese and English test datasets.  
+
+<table><tr><td>Models</td><td>Assistant Turn-taking Acc@K 1/5/10/25 (%)</td><td>Average Assistant Turn-taking Response Time (ms)</td><td>User Turn-taking Acc @K 1/5/10/25 (%)</td><td>Average User Turn-taking Response Time (ms)</td></tr><tr><td>Moshi</td><td>2.9/18.8/38.5/55.1</td><td>553</td><td>0.0/6.2/14.8/45.7</td><td>753</td></tr><tr><td>OmniFlatten</td><td>20.6/53.6/ 66.3/71.7</td><td>193</td><td>10.9/30.9/41.8/51.8</td><td>287</td></tr></table>
+
+Table 4: Assistant Turn-taking and User Turn-taking accuracy at the k-th token (Acc@K) and Response Time.
+
+outputs, with these models occasionally incorporating unclear semantic content in speech endings, adversely affecting scores. When comparing models that produce both speech and text, the performance in the speech modality consistently falls short of that in the text modality, indicating a slight compromise in accuracy despite the ability to synchronize speech with text outputs.
+
+OmniFlatten exhibits superior chat performance on the English test set compared to Moshi, which frequently resorts to counter-questioning the user instead of providing direct responses. In contrast, compared with the 7B parameter LLaMA-Omni, OmniFlatten lagging scores on the English dataset while leading in the Chinese dataset. The proficiency in Chinese for LLaMA-Omni predominantly stems from the speech-text alignment process, despite the absence of Chinese data in its dialogue learning process. Comparatively, against the 9B parameter GLM-Voice model, OmniFlatten's indicators are relatively underperforming, a situation we attribute to potential data leakage during GLM-Voice's training process involving our test set. Additionally, comparing with the Qwen2-0.5B-Instruct results, the inclusion of the speech modality results in a noticeable decline in chat capabilities.
+
+Turn-taking Performance and Runtime Efficiency. To assess the naturalness of OmniFlatten's full-duplex interactions, we evaluate the assistant's ability to respond promptly after the user finishes speaking and to cease speaking when the user attempts to interrupt. In real-world scenar-
+
+ios, timely responses by machines prompt specific human interactive behaviors. For instance, if a machine fails to answer a question quickly, a human might repeat the question. Likewise, if a machine does not stop talking when interrupted, the human may try to interrupt again. Our current testing approach involves using a predefined user speech input while setting a 1.5-second threshold. We consider it a failure of user turn-taking or assistant turn-taking if, exceeded by either, the machine fails to respond (i.e., start or stop answering) within the threshold. Based on this principle, we define the following metrics:
+
+Assistant Turn-taking Acc@k: This metric is defined as whether assistant correctly predicts a non-silence token at the k-th token after the end of a semantically meaningful speech token from user, indicating that assistant starts speaking.
+
+User Turn-taking Acc@k: This metric is defined as whether assistant correctly outputs a silence token at the k-th token after a semantically meaningful speech token is input from user while assistant is speaking. This metric indicates that assistant has successfully responded to the turn-taking attempt from user by stopping assistant's speech and listening.
+
+The evaluation results are shown in Table 4. We compared the response times of OmniFlatten and Moshi during interactions involving user turn-taking and assistant turn-taking scenarios. It was observed that OmniFlatten demonstrates a shorter average response time compared to Moshi in both user and assistant turn-taking contexts. However,
+
+in terms of the user turn-taking metric, although OmniFlatten leads in response time and accuracy, neither model exhibits a particularly high success rate in achieving user turn-taking within 25 tokens.
+
+# 5 Conclusion
+
+In this paper, we introduce an end-to-end full-duplex spoken dialogue model OmniFlatten based on synthesizing full-duplex spoken dialogue data and designing a multi-stage progressive training paradigm for modality alignment and dialogue learning. Our approach offers a straightforward full-duplex modeling scheme as it does not change the architecture of the backbone text-based LLM nor relies on computationally intensive pre-training. Empirical evaluations show that the proposed approach is promising for developing end-to-end models to handle full-duplex interactions.
+
+In future work, we plan to refine the data synthesis pipeline and better simulate the complex interaction patterns, such as user backchannels. Additionally, we will explore more potentials of this modeling scheme by extending to full-duplex interaction involving more modalities, such as vision.
+
+# 6 Limitations
+
+We acknowledge the limitations of this paper. Our work is based on training with the qwen2-0.5B model and utilizes a relatively small scale of dialogue training data. Compared to other similar models that have been scaled up, our model exhibits considerable potential for enhancement in chat capabilities. Particularly, the response speed of our model, especially in scenarios involving user turn-taking, requires further optimization. We have identified certain conflicts between the goal of early stopping by the assistant and some training objectives in speech-text alignment. Moreover, our current model is not yet capable of handling more complex human-machine interaction phenomena such as user back-channel or even assistant back-channel, indicating that the naturalness of human-machine interaction could be further enhanced.
+
+# References
+
+Keyu An, Qian Chen, Chong Deng, Zhihao Du, Changfeng Gao, Zhifu Gao, Yue Gu, Ting He, Hangrui Hu, Kai Hu, Shengpeng Ji, Yabin Li, Zerui Li, Heng Lu, Haoneng Luo, Xiang Lv, Bin Ma, Ziyang Ma, Chongjia Ni, Changhe Song, Jiaqi Shi, Xian Shi, Hao Wang, Wen Wang, Yuxuan Wang,
+
+Zhangyu Xiao, Zhijie Yan, Yexin Yang, Bin Zhang, Qinglin Zhang, Shiliang Zhang, Nan Zhao, and Siqi Zheng. 2024. Funaudiollm: Voice understanding and generation foundation models for natural interaction between humans and llms. Preprint, arXiv:2407.04051.  
+Yunfei Chu, Jin Xu, Qian Yang, Haojie Wei, Xipin Wei, Zhifang Guo, Yichong Leng, Yuanjun Lv, Jinzheng He, Junyang Lin, Chang Zhou, and Jingren Zhou. 2024. Qwen2-audio technical report. Preprint, arXiv:2407.10759.  
+Alexandre Defossez, Laurent Mazaré, Manu Orsini, Amélie Royer, Patrick Pérez, Hervé Jégou, Edouard Grave, and Neil Zeghidour. 2024. Moshi: a speech-text foundation model for real-time dialogue. Technical report, Kyutai.  
+Ning Ding, Yulin Chen, Bokai Xu, Yujia Qin, Shengding Hu, Zhiyuan Liu, Maosong Sun, and Bowen Zhou. 2023. Enhancing chat language models by scaling high-quality instructional conversations. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, EMNLP 2023, Singapore, December 6-10, 2023, pages 3029-3051. Association for Computational Linguistics.  
+Zhihao Du, Qian Chen, Shiliang Zhang, Kai Hu, Heng Lu, Yexin Yang, Hangrui Hu, Siqi Zheng, Yue Gu, Ziyang Ma, Zhifu Gao, and Zhijie Yan. 2024a. Cosyvoice: A scalable multilingual zero-shot text-to-speech synthesizer based on supervised semantic tokens. Preprint, arXiv:2407.05407.  
+Zhihao Du, Jiaming Wang, Qian Chen, Yunfei Chu, Zhifu Gao, Zerui Li, Kai Hu, Xiaohuan Zhou, Jin Xu, Ziyang Ma, Wen Wang, Siqi Zheng, Chang Zhou, Zhijie Yan, and Shiliang Zhang. 2024b. Lauragpt: Listen, attend, understand, and regenerate audio with gpt. Preprint, arXiv:2310.04673.  
+Qingkai Fang, Shoutao Guo, Yan Zhou, Zhengrui Ma, Shaolei Zhang, and Yang Feng. 2024. Llama-omni: Seamless speech interaction with large language models. Preprint, arXiv:2409.06666.  
+Chaoyou Fu, Haojia Lin, Zuwei Long, Yunhang Shen, Meng Zhao, Yifan Zhang, Shaoqi Dong, Xiong Wang, Di Yin, Long Ma, Xiawu Zheng, Ran He, Rongrong Ji, Yunsheng Wu, Caifeng Shan, and Xing Sun. 2024. Vita: Towards open-source interactive omni multimodal llm. Preprint, arXiv:2408.05211.  
+Zhifu Gao, Zerui Li, Jiaming Wang, Haoneng Luo, Xian Shi, Mengzhe Chen, Yabin Li, Lingyun Zuo, Zhihao Du, and Shiliang Zhang. 2023. Funasr: A fundamental end-to-end speech recognition toolkit. In 24th Annual Conference of the International Speech Communication Association, Interspeech 2023, Dublin, Ireland, August 20-24, 2023, pages 1593-1597. ISCA.  
+François Hernandez, Vincent Nguyen, Sahar Ghannay, Natalia A. Tomashenko, and Yannick Esteve. 2018. TED-LIUM 3: Twice as much data and corpus repartition for experiments on speaker adaptation. In
+
+Speech and Computer - 20th International Conference, SPECOM 2018, Leipzig, Germany, September 18-22, 2018, Proceedings, volume 11096 of Lecture Notes in Computer Science, pages 198-208. Springer.  
+Yunjie Ji, Yan Gong, Yong Deng, Yiping Peng, Qiang Niu, Baochang Ma, and Xiangang Li. 2023. Towards better instruction following language models for Chinese: Investigating the impact of training data and evaluation. Preprint, arXiv:2304.07854.  
+Jungil Kong, Jaehyeon Kim, and Jaekyoung Bae. 2020. Hifi-gan: Generative adversarial networks for efficient and high fidelity speech synthesis. In NeurIPS.  
+Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. 2023. Flow matching for generative modeling. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. Open-Review.net.  
+Ziyang Ma, Yakun Song, Chenpeng Du, Jian Cong, Zhuo Chen, Yuping Wang, Yuxuan Wang, and Xie Chen. 2024. Language model can listen while speaking. CoRR, abs/2408.02622.  
+Tu Anh Nguyen, Eugene Kharitonov, Jade Copet, Yossi Adi, Wei-Ning Hsu, Ali Elkahky, Paden Tomasello, Robin Algayres, Benoit Sagot, Abdelrahman Mohamed, and Emmanuel Dupoux. 2023. Generative spoken dialogue language modeling. Trans. Assoc. Comput. Linguistics, 11:250-266.  
+Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. 2015. Librispeech: An asr corpus based on public domain audio books. In 2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 5206-5210.  
+Baolin Peng, Chunyuan Li, Pengcheng He, Michel Galley, and Jianfeng Gao. 2023. Instruction tuning with GPT-4. CoRR, abs/2304.03277.  
+Vineel Pratap, Qiantong Xu, Anuroop Sriram, Gabriel Synnaeve, and Ronan Collobert. 2020. Mls: A large-scale multilingual dataset for speech research. ArXiv, abs/2012.03411.  
+Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. 2023. Robust speech recognition via large-scale weak supervision. In International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA, volume 202 of Proceedings of Machine Learning Research, pages 28492-28518. PMLR.  
+Takaaki Saeki, Detai Xin, Wataru Nakata, Tomoki Koriyama, Shinnosuke Takamichi, and Hiroshi Saruwatari. 2022. UTMOS: utokyo-sarulab system for voicemos challenge 2022. CoRR, abs/2204.02152.  
+David Snyder, Guoguo Chen, and Daniel Povey. 2015. MUSAN: A Music, Speech, and Noise Corpus. ArXiv:1510.08484v1.
+
+Tianxiang Sun, Xiaotian Zhang, Zhengfu He, Peng Li, Qinyuan Cheng, Xiangyang Liu, Hang Yan, Yunfan Shao, Qiong Tang, Shiduo Zhang, Xingjian Zhao, Ke Chen, Yining Zheng, Zhejian Zhou, Ruixiao Li, Jun Zhan, Yunhua Zhou, Linyang Li, Xiaogui Yang, Lingling Wu, Zhangyue Yin, Xuanjing Huang, YuGang Jiang, and Xipeng Qiu. 2024. Moss: An open conversational large language model. Machine Intelligence Research, 21(5):888-905.  
+Changli Tang, Wenyi Yu, Guangzhi Sun, Xianzhao Chen, Tian Tan, Wei Li, Lu Lu, Zejun Ma, and Chao Zhang. 2024. Salmonn: Towards generic hearing abilities for large language models. Preprint, arXiv:2310.13289.  
+Alexander Tong, Kilian Fatras, Nikolay Malkin, Guillaume Huguet, Yanlei Zhang, Jarrid Rector-Brooks, Guy Wolf, and Yoshua Bengio. 2024. Improving and generalizing flow-based generative models with minibatch optimal transport. Trans. Mach. Learn. Res., 2024.  
+Bandhav Veluri, Benjamin N Peloquin, Bokai Yu, Hongyu Gong, and Shyamnath Gollakota. 2024. Beyond turn-based interfaces: Synchronous llms as full-duplex dialogue agents. Preprint, arXiv:2409.15594.  
+Changhan Wang, Morgane Riviere, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Miguel Pino, and Emmanuel Dupoux. 2021. Voxpopuli: A large-scale multilingual speech corpus for representation learning, semi-supervised learning and interpretation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, ACL/IJCNLP 2021, (Volume 1: Long Papers), Virtual Event, August 1-6, 2021, pages 993-1003. Association for Computational Linguistics.  
+Zhifei Xie and Changqiao Wu. 2024a. Mini-omni: Language models can hear, talk while thinking in streaming. Preprint, arXiv:2408.16725.  
+Zhifei Xie and Changqiao Wu. 2024b. Mini-omni2: Towards open-source gpt-4o with vision, speech and duplex capabilities. Preprint, arXiv:2410.11190.  
+An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jianxin Yang, Jin Xu, Jingren Zhou, Jinze Bai, Jinzheng He, Junyang Lin, Kai Dang, Keming Lu, Keqin Chen, Kexin Yang, Mei Li, Mingfeng Xue, Na Ni, Pei Zhang, Peng Wang, Ru Peng, Rui Men, Ruize Gao, Runji Lin, Shijie Wang, Shuai Bai, Sinan Tan, Tianhang Zhu, Tianhao Li, Tianyu Liu, Wenbin Ge, Xiaodong Deng, Xiaohuan Zhou, Xingzhang Ren, Xinyu Zhang, Xipin Wei, Xuancheng Ren, Xuejing Liu, Yang Fan, Yang Yao, Yichang Zhang, Yu Wan, Yunfei Chu, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, Zhifang Guo, and Zhihao Fan. 2024. Qwen2 technical report.
+
+Xin Xu Shaoji Zhang Ming Li Yao Shi, Hui Bu. 2015. Aishell-3: A multi-speaker mandarin ts corpus and the baselines.
+
+Heiga Zen, Viet Dang, Rob Clark, Yu Zhang, Ron J. Weiss, Ye Jia, Zhifeng Chen, and Yonghui Wu. 2019. Libritts: A corpus derived from librispeech for text-to-speech. In 20th Annual Conference of the International Speech Communication Association, Interspeech 2019, Graz, Austria, September 15-19, 2019, pages 1526-1530. ISCA.
+
+Aohan Zeng, Zhengxiao Du, Mingdao Liu, Kedong Wang, Shengmin Jiang, Lei Zhao, Yuxiao Dong, and Jie Tang. 2024. Glm-4-voice: Towards intelligent and human-like end-to-end spoken chatbot. Preprint, arXiv:2412.02612.
+
+Binbin Zhang, Hang Lv, Pengcheng Guo, Qijie Shao, Chao Yang, Lei Xie, Xin Xu, Hui Bu, Xiaoyu Chen, Chenchen Zeng, Di Wu, and Zhendong Peng. 2022. Wenetspeech: A  $10000+$  hours multi-domain mandarin corpus for speech recognition. In International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE.
+
+Dong Zhang, Shimin Li, Xin Zhang, Jun Zhan, Pengyu Wang, Yaqian Zhou, and Xipeng Qiu. 2023. Speechgpt: Empowering large language models with intrinsic cross-modal conversational abilities. Preprint, arXiv:2305.11000.
+
+Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. 2023a. Judging llm-as-a-judge with mt-bench and chatbot arena. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023.
+
+Siqi Zheng, Luyao Cheng, Yafeng Chen, Hui Wang, and Qian Chen. 2023b. 3d-speaker: A large-scale multi-device, multi-distance, and multi-dialect corpus for speech representation disentanglement. *Preprint*, arXiv:2306.15354.
+
+# A Prompt LLM for Dialogue Quality Evaluation
+
+# Auto-Evaluation Prompt
+
+Please rate the given dialogue context and response from the voice dialogue system based on the following criteria (1-10 points), and provide a brief evaluation:
+
+1. Relevance: Is the response relevant to the query? Is the content related?  
+2. Accuracy: Does the response correctly address the user's query and provide accurate information?  
+3. Completeness: Does the response comprehensively cover all aspects of the query?  
+4. Conversational Nature: Is the response easy to understand, concise, clear, and fluent?
+
+Context: context
+
+Response: response
+
+Output in JSON format:
+
+{
+
+"Strengths": "Positive aspects of the response",
+
+"Weaknesses": "Negative aspects of the response",
+
+"Overall Evaluation": "Overall assessment of the response",
+
+"Total Score (out of 10, directly provide the score): ""   
+}

ACL/2025/OmniFlatten_ An End-to-end GPT Model for Seamless Voice Conversation/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6d8a818839e06577f6894987801a17b4694343d4b87c518862b33018906a072d
+size 294050

ACL/2025/OmniFlatten_ An End-to-end GPT Model for Seamless Voice Conversation/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a8ffa053a0b0934f218b45bf7c43200456eb8b3e9ef1bf14893469d8b8d96239
+size 274222

ACL/2025/On Generalization across Measurement Systems_ LLMs Entail More Test-Time Compute for Underrepresented Cultures/298b445b-7961-4296-870d-cc3c02c89093_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0cc13b8e8f195ae24284c05daf5326fa2763db9f655302545f37c7f26bdb79b9
+size 108062

ACL/2025/On Generalization across Measurement Systems_ LLMs Entail More Test-Time Compute for Underrepresented Cultures/298b445b-7961-4296-870d-cc3c02c89093_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:814297115616c887c0aeb7fded44acb81b2b45e475d42807d45d06a3a4f3e502
+size 125378

ACL/2025/On Generalization across Measurement Systems_ LLMs Entail More Test-Time Compute for Underrepresented Cultures/298b445b-7961-4296-870d-cc3c02c89093_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6e975a4d2deb9809b5a15373683a10933b4600ab4beb870ad2ce426120d9d74f
+size 819243

ACL/2025/On Generalization across Measurement Systems_ LLMs Entail More Test-Time Compute for Underrepresented Cultures/full.md

@@ -0,0 +1,370 @@
+# On Generalization across Measurement Systems: LLMs Entail More Test-Time Compute for Underrepresented Cultures
+
+Minh Duc Bui $^{\diamond}$  Kyung Eun Park $^{\clubsuit}$  Goran Glavas $^{\diamond}$  Fabian David Schmidt $^{\diamond}$  Katharina von der Wense $^{\diamond\clubsuit}$
+
+$\nabla$  Johannes Gutenberg University Mainz, Germany  $\clubsuit$ University of Mannheim, Germany  
+ $\diamond$ Center For Artificial Intelligence and Data Science, University of Wurzburg, Germany  $\spadesuit$ University of Colorado Boulder, USA
+
+{minhducbui, k.vonderwense}@uni-mainz.de kyung.eun.park@uni-mannheim.de {fabian.schmidt, goran.glavas}@uni-wuerzburg.de
+
+# Abstract
+
+Measurement systems (e.g., currencies) differ across cultures, but the conversions between them are well defined so that humans can state facts using any measurement system of their choice. Being available to users from diverse cultural backgrounds, large language models (LLMs) should also be able to provide accurate information irrespective of the measurement system at hand. Using newly compiled datasets we test if this is the case for seven open-source LLMs, addressing three key research questions: (RQ1) What is the default system used by LLMs for each type of measurement? (RQ2) Do LLMs' answers and their accuracy vary across different measurement systems? (RQ3) Can LLMs mitigate potential challenges w.r.t. underrepresented systems via reasoning? Our findings show that LLMs default to the measurement system predominantly used in the data. Additionally, we observe considerable instability and variance in performance across different measurement systems. While this instability can in part be mitigated by employing reasoning methods such as chain-of-thought (CoT), this implies longer responses and thereby significantly increases test-time compute (and inference costs), marginalizing users from cultural backgrounds that use underrepresented measurement systems.
+
+# 1 Introduction
+
+The user base of large language models (LLMs) expands to people from a wide range of cultural backgrounds, with each person bringing distinct perspectives to their interactions with these ubiquitous systems. Intuitively, users expect that data can be correctly presented in ways that align with their cultural context, including measurement systems that they are used to. For instance, individuals accustomed to the metric system might find it challenging to interpret imperial units.
+
+Prior work shows that LLMs encode factual information, including plethora of measurement val
+
+![](images/8f95de38e7af7d9e6a3d34ef5723926dada492b958ae3438055810e27f093b84.jpg)  
+Figure 1: We test whether LLMs can translate facts across different measurement systems and show that LLMs often struggle to provide accurate information for cultural contexts underrepresented in training data.
+
+ues, within their parameters (Petroni et al., 2019; Geva et al., 2023; Allen-Zhu and Li, 2024, inter alia). However, while an LLM may accurately present facts within one cultural context (presumably the one that dominated their respective training corpora), an inclusive system should be able to do so in all contexts. By failing to do so, LLMs would disadvantage users from cultural backgrounds that are less represented in their training data. This raises a critical, yet underinvestigated question: can LLMs effectively generalize across socio-cultural contexts, offering same levels of factual correctness irrespective of the concrete culturally-bound factors such as measurement systems? Or do they exhibit biases that hinder sociocultural adaptability, preventing them from equitably providing accurate information to all users?
+
+In this work, we demonstrate that LLMs are less reliable when the users' measurement system differs from the dominant system of the LLM. We find that LLMs typically default to the measurement that fits the cultural context of the data, e.g., reporting fiscal values in USD as well as length and
+
+weight in the metric or imperial system, depending on the context of the data. For instance LLMs report food prices of US goods in USD normalized per pounds. The main issue arises when users naively prompt the models with systems that differ from the dominant one, as LLMs often fail to successfully translate values from one measurement into another (e.g., monetary amounts from USD to SDG). While advanced prompting techniques such as chain-of-thought reasoning largely mitigate these conversion errors (e.g., from  $72\%$  to almost no difference for currency for the largest LLMs in our evaluation), we need to note that this improvement comes at a substantial cost. Under CoT, models produce much longer, more verbose responses as they triangulate the conversion process step-by-step. For example, obtaining the correct value for the GDP of a country in Eritrean Nakfa can increase the inference cost by up to  $300\%^{1}$  compared to requesting the same value in US Dollars. Such issues have severe implications for users from marginalized cultural backgrounds, who are meant to benefit the most from equitable access to language technology. When LLMs misrepresent or inaccurately convert critical measurements (e.g., wrongly measuring monetary value of consumer goods), they alienate users by failing to align with their cultural context. Moreover, as these users frequently come from economically disadvantaged regions, they are disproportionately affected by the increased inference costs associated with extended model responses.
+
+We address the following research questions in this work (together with respective contributions).
+
+# (RQ1) What measurement do LLMs default to?
+
+We curate datasets that span diverse cultural contexts, including fiscal data, food prices, and city distances across different countries and regions. We find that LLMs default to the measurement that fits the cultural context of the data. When queried on distances between German and US cities, they provide estimates in the metric and imperial units, respectively. Likewise, LLMs express fiscal data in USD, the common convention for financial data.
+
+# (RQ2) Do LLMs' answers and their accuracy vary across measurement systems?
+
+We test whether varying measurement systems affects factual accuracy of answers provided by LLMs (as illustrated in Figure 1). We find that
+
+LLMs exhibit significant variability in behavior depending on the measurement system embedded in the prompt. Models perform best when the query aligns with the measurement system that fits the cultural context of the data (e.g., provided by a location or entity mentions). For instance, models show a strong decline in accuracy when asked about food prices in pounds for a country that uses kilograms.
+
+# (RQ3) Can LLMs mitigate these challenges with reasoning?
+
+We find that LLMs often struggle to generate accurate facts across different measurement systems. Nevertheless, very large LLMs (70B parameters) can successfully perform sequential single-hop inferences, starting with the default measurement and then converting the value to the target unit. This exemplifies that LLMs necessitate additional guidance to reliably triangulate to the correct value in the target measurement. Explicitly prompting LLMs to reason across measurement systems similarly stabilizes performance. Yet, such reasoning remains anchored in the default measurement, increasing inference costs for queries reflecting non-default cultural contexts.
+
+# 2 Measurement Systems & Datasets
+
+# 2.1 Measurement Systems
+
+Each dataset is rooted in a specific cultural context. For instance, maps often use local units of measurement, international fiscal data such as GDP is typically reported in USD, and food prices are commonly expressed in local currencies. While previous studies have explored the geographical biases in LLMs' factual retrieval (Manvi et al., 2024; Moayeri et al., 2024), there has been limited investigation into whether these models can effectively adapt facts to diverse socio-cultural contexts. For example, can a model adapt fiscal data to a different currency or report city distances using an alternative length system? To address this gap, we focus on three measurement systems: currency, length systems, and weight systems. In Appendix A.1, we describe each system in depth and report how often each remains in use nowadays.
+
+# 2.2 Datasets
+
+Table 1 provides an overview of the measurement systems employed in each dataset, including details on their total sizes and representative examples. This dataset is released under the
+
+<table><tr><td></td><td>Fiscal Data</td><td>City Distances</td><td>Food Prices per Weight</td></tr><tr><td>Measurement System Units</td><td>CurrencyUSD, EUR, VND, KRW, CNY, JPY, ...</td><td>Length SystemsKilometer (Metric), Mile (Imperial), Cape Foot (South African), Wa (Thai), Li (Chinese), ...</td><td>Weight SystemsKilogram (Metric), Pound (Imperial) Jin (Chinese), Geun (Korean), Catty/Kati (Malaysian), ...</td></tr><tr><td>Number of Systems</td><td>112</td><td>10</td><td>10</td></tr><tr><td>Total Samples (Facts × Systems)</td><td>16576</td><td>31200 (Kilometer), 7800 (Mile)</td><td>24130 (Kilogram), 3690 (Pound)</td></tr><tr><td>Number of Facts</td><td>148</td><td>3120 (Kilometer), 780 (Mile)</td><td>2413 (Kilogram), 369 (Pound)</td></tr><tr><td>Example Question</td><td>What is GDP per capita for Australia in 2021 in JPY?</td><td>What is the distance between New York and Los Angeles in Li (Chinese)?</td><td>What is the price of one Korean Geun of steak (beef) in Nigeria, in Nigerian Naira (NGN)?</td></tr></table>
+
+Table 1: Measurement System Datasets: Measurement system and units, as well as sample size by dataset type. For further information about each measurement system, see Appendix Table 5.
+
+CC BY-NC-ND 4.0 license. For complete details on usage terms and proper attribution, please refer to Appendix A.2. Code and dataset are available at https://github.com/MinhDucBui/MeasurementSystemBias.
+
+Fiscal Data. To analyze currency systems, we collect international fiscal data focused on GDP per capita. We obtain the figures for 148 countries from the World Bank (The World Bank, 2024) and specifically use 2021 data to align with the models' knowledge cutoff, while at the same time minimizing recency biases (Moayeri et al., 2024). We obtain the average monthly exchange rates for 2021 from the International Monetary Fund to convert these fiscal values into other currencies (International Monetary Fund, 2024). To mitigate the impact of exchange rate fluctuations, we select the most favorable rate within that year, ensuring optimal performance. Note that fiscal data is typically reported in USD (Krugman, 1984).
+
+Food Prices per Weight: Kilogram System. We differentiate between countries that use the "Kilogram" (metric) system and the "Pound" (imperial) system. In countries that use the "Kilogram" system, we collect food commodity prices for items such as maize, rice, beans, and sugar reported per kilogram (e.g., price of  $1\mathrm{kg}$  of rice), expressed in the respective local currency. This data stems from World Food Program via HDX (World Food Program, 2025), covers 76 countries, and spans from 2010 to 2021.
+
+Food Prices per Weight: Pound System. We obtain food price data for United States from the Bureau of Labor Statistics (U.S. Bureau of Labor Statistics, 2025). This dataset includes retail prices for a wide range of items, including fruits, meats,
+
+and other commodities, expressed in pound (e.g., the price of 1 pound of chicken).
+
+City Distances: Kilometer System. We begin by selecting four countries in which the "Kilometer" (metric) system is the primary measurement system: Germany, Russia, China and Japan. For each country, we randomly select 40 cities from among its 80 largest cities. We then compute the pairwise distances between all selected cities within each country, resulting in 3,120 distance measurements in total. The shortest straight-line distances between cities were calculated using the Haversine formula, with city coordinates sourced from SimpleMaps (2025).
+
+City Distances: Mile System. For the United States, where the "Mile" (imperial) system applies, we likewise compile a list of 80 largest cities by population and calculate the shortest road distances as for the "Kilometer" system.
+
+# 3 Stability across Measurement Systems
+
+We begin our investigation of RQ1 by analyzing what measurement systems LLMs default to when prompted without any instruction on the use of measurement. Next, we investigate RQ2 by examining whether LLM behavior shifts when instructed to use a measurement system different from its default and whether this affects performance.
+
+# 3.1 Experimental Setup
+
+Models. We test instruct-finetuned LLMs from different model families and of varying sizes. Specifically, we probe Qwen2.5 (72B, 7B) (Qwen et al., 2025), Llama 3.3 (72B), Llama 3.1 (72B, 8B) (Grattafori et al., 2024) and Aya Expanse (32B, 8B) (Dang et al., 2024). We group the models by
+
+<table><tr><td>System</td><td>Large</td><td>Small</td><td>Avg.</td></tr><tr><td colspan="4">Currency (Fiscal Data)</td></tr><tr><td>Local Currency</td><td>18%</td><td>16%</td><td>17%</td></tr><tr><td>Other = USD</td><td>76%</td><td>82%</td><td>80%</td></tr><tr><td>= Rest</td><td>6%</td><td>2%</td><td>3%</td></tr><tr><td colspan="4">Weight Systems (Food Prices)</td></tr><tr><td>Data: Kilogram</td><td></td><td></td><td></td></tr><tr><td>Kilogram</td><td>100%</td><td>100%</td><td>100%</td></tr><tr><td>Pound</td><td>0%</td><td>0%</td><td>0%</td></tr><tr><td>Geun, Jin, ...</td><td>0%</td><td>0%</td><td>0%</td></tr><tr><td>Data: Pound</td><td></td><td></td><td></td></tr><tr><td>Kilogram</td><td>0%</td><td>0%</td><td>0%</td></tr><tr><td>Pound</td><td>100%</td><td>100%</td><td>100%</td></tr><tr><td>Geun, Jin, ...</td><td>0%</td><td>0%</td><td>0%</td></tr><tr><td colspan="4">Length Systems (City Distances)</td></tr><tr><td>Data: Kilometer</td><td></td><td></td><td></td></tr><tr><td>Kilometer</td><td>100%</td><td>100%</td><td>100%</td></tr><tr><td>Mile</td><td>0%</td><td>0%</td><td>0%</td></tr><tr><td>Cape Foot, Wa, ...</td><td>0%</td><td>0%</td><td>0%</td></tr><tr><td>Data: Mile</td><td></td><td></td><td></td></tr><tr><td>Kilometer</td><td>37%</td><td>22%</td><td>27%</td></tr><tr><td>Mile</td><td>63%</td><td>78%</td><td>73%</td></tr><tr><td>Cape Foot, Wa, ...</td><td>0%</td><td>0%</td><td>0%</td></tr></table>
+
+Table 2: LLMs' Default Measurement Systems: We report the total percentage of measurement systems that LLMs default to when given the freedom to choose.
+
+their size into 'Large'  $(>=70\mathrm{B})$  and 'Small'  $(<  70\mathrm{B})$  For more information, see Appendix B.2.
+
+**Prompts.** To first isolate the effect of different measurement systems, we use English prompts in our experiments, ruling out the language of the prompt as a confounding factor. In §3.3, we also explore multilingual prompts. For robustness, we generate three prompt variants for each task. In each variant, the model is instructed to adhere to a specific measurement system, as illustrated by the following example:
+
+Estimate the GDP per capita for Germany in the year 2021, expressed in the currency ERN. Your answer must only include the GDP per capita [...]
+
+We refer to Appendix B.1 for further details. Additionally, we specify the original name and country of origin for certain length and weight systems.
+
+# 3.2 RQ1: Default Systems
+
+We first analyze which measurement system the model defaults to. To that end, we let the model generate its response without the instruction to output the fact in a specific measurement.
+
+Results. Table 2 summarizes the findings on default systems. For fiscal data, LLMs predominantly
+
+default to using USD (80% on average), and almost exclusively rely on either USD or local currency (17% on average). For weight measurements, LLMs strictly employ either the imperial or metric system, in line with the data context. In the case of length data, distances in Germany, Japan, Russia and China are reported solely in the metric system, whereas, for city distances in the USA, 73% are expressed in the imperial system.
+
+Discussion. Our observations suggest that LLMs tend to adopt measurement systems that mirror the conventions present in the data, often defaulting to Western standards, such as using USD for fiscal figures. This trend seems to reflect the broader influence of the West, emerging not through any direct global mandate but through market dynamics shaped by Western forces $^{2}$  (Krugman, 1984). Interestingly, Large and Small LLMs exhibit very similar behavior, with US city distances as the only (partial) exception: Small LLMs default more consistently  $(+15\%)$  to the American cultural context, i.e., the imperial system.
+
+# 3.3 RQ2: Disparity & Performance across Measurement Systems
+
+We evaluate the stability of LLMs' responses across measurement systems by explicitly instructing them to produce outputs in a specified system. For each dataset, the default system is the one which LLMs used most frequently in §3.2 (Table 2). We then use the performance with the default system as a baseline for the alternative systems.
+
+Evaluation Metric. We evaluate a model's performance by comparing its predictions to the ground truth. Because the magnitude of answers varies considerably, we resort to an inverse mean absolute percentage deviation (MAPD). To mitigate the impact of extreme predictions—which LLMs may occasionally produce—we cap the deviation at a maximum of  $100\%$ . The final metric is defined as follows:
+
+$$
+A P D _ {i} ^ {k} = \min \left(\left| \frac {p (x _ {i} , f ^ {k}) ^ {z} - y _ {i} ^ {z}}{y _ {i} ^ {z}} \right|, 1\right),
+$$
+
+$$
+\mathrm {M A P D} ^ {k} = 1 0 0 \times \left(1 - \frac {1}{N} \sum_ {i = 1} ^ {N} A P D _ {i} ^ {k}\right).
+$$
+
+![](images/31fe2fc4525384f962a306f7f4e744c349fca792f102144b08a584cda190e988.jpg)
+
+![](images/065e9541d2cc3a2cfd92d8d208151776729faea281e40a7a47622afb0e8f113e.jpg)
+
+![](images/53d6a5a7605d13281519f4072cc3af8a4e272a8d079eb4e2909434da43a61c94.jpg)
+
+![](images/dcdae137fc1086d062f21d44fb775d8c749f3245766eaa1a298aa4694154e127.jpg)  
+Figure 2: Disparity and Performance of LLMs Across Different Measurement Systems: We report the performance in terms of MAPD (cf. §3.3). The best-performing system (higher is better) for each model is highlighted with bold points. Fully filled points indicate statistically significant improvement over the second-best system, whereas half-filled points do not. We report the raw numbers and their significance in Appendix Table 6.
+
+![](images/93ca3b2776e8c5cdf484be14a2252d0c86cf4ebedda4151fad4f4a61a1af0db5.jpg)
+
+Here,  $p(x_{i},f^{k})$  is the model's prediction for sample  $i$  using system  $k$ , and  $p(x_{i},f^{k})^{z}$  represents its conversion to measurement system  $z$ . The ground truth value for sample  $i$ , also expressed in system  $z$ , is denoted by  $y_{i}^{z}$ . We note that higher values indicate better performance. An inverse MAPD of 76.0 indicates approximately that the model's estimates deviate from the ground truth by  $24\%$ .
+
+We test whether the behavior of two systems differs significantly by comparing their paired absolute percentage deviation distributions over samples. We employ the non-parametric Wilcoxon Signed-Rank Test (Wilcoxon, 1945), which establishes whether the median difference between the paired values significantly deviates from zero. We use a high significance threshold of  $p = 0.001$ .
+
+Results. Table 2 summarizes our findings. On average across all models, the default system consistently delivers the best performance. Among 35 model-dataset combinations, it outperforms all others in 28 cases, with statistically significant gains over the second-best system in 22 of those. For example, in the fiscal dataset, using USD yields
+
+the highest performance for every single model (average MAPD of  $63\%$ ). Similarly, in the weight and length system dataset, models overwhelmingly yield optimal results with the default system. To illustrate, Llama 3.3 70B achieves a  $57\%$  MAPD with the default "Kilogram" system, but its performance drops significantly to  $29\%$  when using the "Pound" system. In the length dataset, Llama 3.3 70B achieves a  $75\%$  MAPD on cities with the default "Kilometer" system and shows a 4-point performance drop when evaluated with the "Mile" system. Only for US cities, we observe a smaller average difference from the "Kilometer" system and with Qwen2.5 7B being the only model to show a statistically significant improvement when using this alternative system. While other models also perform better with certain alternative systems, their improvements are not statistically significant. Finally, we observe uneven performance drops across alternative measurement systems, as reflected in the drop of average scores.
+
+Discussion. In sum, we find that current LLMs are not stable across different measurement sys
+
+![](images/68e7ee30c78d07336ec46d4ab498ed75d651acd51295f6a0459c36bc954bb712.jpg)  
+Figure 3: Currency Performance Mapped and Categorized by Income Groups: We visualize the performance for all currencies—averaged across large models (over 70B parameters)—on a world map, with countries categorized by income levels.
+
+tems; LLMs yield the best performance when used with their default system and generally suffer significant and pronounced drops in performance for alternative systems. Such issues have severe implications for users from marginalized cultural backgrounds, who stand to benefit the most from equitable access to language technology. When LLMs misrepresent or inaccurately convert measurements, e.g., wrongly state the value of consumer goods in local currency, they alienate users by failing to correctly adapt to their cultural context.
+
+Analysis: Currency Bias. We visualize currency performance of our biggest LLMs  $(>=70\mathrm{B})$  on a world map, associating each country with its respective currency. Additionally, we group countries by income levels provided by the World Bank (Fantom and Serajuddin, 2016). We immediately observe obvious disparities across income groups. LLMs achieve an average performance of  $43\%$  with currencies from high-income countries, which is an impressive  $34\%$  better than with currencies from low-income countries. This constitutes a clear bias that favors the richest (i.e., their currencies) and affects the poorest. As already highlighted in §3.3, using non-default systems adversely affects LLMs' performance: this analysis reveals that this effect disproportionately disadvantages low-income countries.
+
+Analysis: Multilingual Setting. We further examine whether language affects our findings by aligning the prompt language with the measurement system used. For instance, when using the Korean currency, we analyze the performance gap between the Korean currency and USD (default measurement system) under Korean prompt language. For each dataset, we select two languages from among Turkish (TR), Korean (KR), Chinese (ZH), and Japanese
+
+<table><tr><td>Setting</td><td>Llama 3.3
+70B</td><td>Qwen2.5
+72B</td><td>Aya
+32B</td></tr><tr><td colspan="4">Currency (Fiscal Data)</td></tr><tr><td>Default=USD</td><td></td><td></td><td></td></tr><tr><td>ΔKR (KRW)</td><td>-37.89*</td><td>-27.43*</td><td>-36.36*</td></tr><tr><td>ΔTR (TRY)</td><td>-57.01*</td><td>-42.22*</td><td>-49.13*</td></tr><tr><td colspan="4">Weight System (Food Prices)</td></tr><tr><td>Default=Kilogram</td><td></td><td></td><td></td></tr><tr><td>ΔKR (Geun)</td><td>-21.99*</td><td>-31.12*</td><td>-22.11*</td></tr><tr><td>ΔZH (Jin)</td><td>-25.39*</td><td>-26.79*</td><td>-11.33*</td></tr><tr><td>Default=Pound</td><td></td><td></td><td></td></tr><tr><td>ΔKR (Geun)</td><td>-3.70</td><td>-12.54*</td><td>-39.14*</td></tr><tr><td>ΔZH (Jin)</td><td>-1.31</td><td>-20.25*</td><td>-2.29</td></tr><tr><td colspan="4">Length System (City Distances)</td></tr><tr><td>Default=Kilometer</td><td></td><td></td><td></td></tr><tr><td>ΔJA (Ken)</td><td>-80.09*</td><td>-75.71*</td><td>-73.90*</td></tr><tr><td>ΔZH (Li)</td><td>-20.19*</td><td>-13.18*</td><td>-15.03*</td></tr><tr><td>Default=Mile</td><td></td><td></td><td></td></tr><tr><td>ΔJA (Ken)</td><td>-89.68*</td><td>-86.97*</td><td>-84.98*</td></tr><tr><td>ΔZH (Li)</td><td>-30.33*</td><td>-35.26*</td><td>-35.45*</td></tr></table>
+
+Table 3: Aligning Language with Measurement System: We compare performance between the default system and the system aligned with the prompt language (e.g., KRW vs. USD for Korean prompts). The language is shown as a subscript, and the aligned system in brackets. Red cells mark significant differences.
+
+(JA)—representing high- to medium-resource languages. For each language, a native speaker reviews our prompts to ensure correctness. We report the results in Table 3.
+
+In nearly all cases, we observe a significant performance drop. For example, even when using the Korean currency with Korean prompt language, Llama 3.3 70B still experiences a significant decrease of 38 points. The only cases without a significant drop are Llama 3.3 70B in the weight task using the "Pound" system and Aya 32B when evaluating the Chinese "Jin" unit, also under the "Pound" system. We conclude that even when the language
+
+is aligned, significant drops persist in nearly all cases, highlighting the significant instability of measurement systems independent of language in almost all cases.
+
+![](images/2fe38a763b8a1833338fd31a62972cdf7034f3367198aa733ffe70649adaf6b5.jpg)  
+Figure 4: Visualization of Multi-Hop Reasoning: Answering w.r.t. an alternative measurement system requires multi-hop reasoning, consisting of the following steps: (i) retrieving the answer in the default system, (ii) retrieving the conversion rate between the default and alternative system, and (iii) combine the results of (i) and (ii) with arithmetic reasoning.
+
+# 4 RQ3: Multi-Hop Reasoning
+
+We first explain how answering questions w.r.t. alternative measurement systems represents a multihop reasoning problem and then explore whether the LLMs' performance changes with (i) a sequence of prompts that aim at individual reasoning steps and (ii) with chain-of-thought prompts, i.e., by explicitly asking the LLMs to provide the explanation for their final answer.
+
+Multi-Hop Questions. We adopt the definition and notation from Mavi et al. (2024): A multi-hop question (MHQ) is a question that cannot be answered using a single document or passage (context); instead, it requires combining information from multiple sources. For example, a question like "What is the capital of the country whose currency has the highest exchange rate against the US dollar?" necessitates gathering and synthesizing information from several contexts.
+
+Formally, let  $\mathbb{C}$  denote the set of all contexts,  $\mathcal{Q}$  the set of all questions, and  $\mathcal{A}$  the set of all possible answers. For a question  $q\in \mathcal{Q}$  and a subset of related contexts  $C\subseteq \mathbb{C}$ , the task is to approximate a function
+
+$$
+f: \mathcal {Q} \times 2 ^ {\mathbb {C}} \rightarrow \mathcal {A} \cup \{\Phi \},
+$$
+
+such that
+
+$$
+f (q, C) = \left\{ \begin{array}{l l} a \in \mathcal {A} & \text {i f t h e r e e x i s t s a s u b s e t} \\ & P _ {q} = \{p _ {1}, \ldots , p _ {k} \} \subseteq C, \\ & \text {w i t h k > 1 , s u c h} \\ & \text {t h a t} P _ {q} \models (a \text {a n s w e r s} q), \\ \Phi & \text {o t h e r w i s e .} \end{array} \right.
+$$
+
+where  $\models$  denotes entailment and  $\Phi$  is returned when  $q$  is unanswerable using  $C$ . The requirement that  $k > 1$  ensures that the question is not solvable in a single hop (i.e., with  $k = 1$ , the task reduces to traditional QA).
+
+When we ask questions of the form "What is  $< \text{Fact} >$  in  $< \text{System} >$ ?", where  $< \text{Fact} >$  represents information inherently associated with a particular measurement system (e.g., GDP is intrinsically tied to currency) and the system is not the default one for the model, we are effectively asking the model to solve a multi-hop reasoning task consisting of the following steps:
+
+1. Source Fact Retrieval: Extract the fact from its original measurement system (the source system,  $S_{\mathrm{src}}$ ). This constitutes a single-hop operation using one context,  $p_1$ .  
+2. Conversion Retrieval: Obtain the information required to convert the fact from the source system  $S_{\mathrm{src}}$  to the target system  $S_{\mathrm{trg}}$ . This is another single-hop operation that relies on a different context,  $p_2$ .  
+3. Arithmetic Reasoning: Apply the conversion information to the retrieved fact. This requires arithmetic reasoning and may require one or more additional contexts  $\{p_3,\ldots \}$
+
+In this scenario, the model must integrate multiple contexts  $\{p1, p2, p3, \ldots\}$  to arrive at an answer, which aligns with our multi-hop definition. Figure 4 provides a visual overview of this process.
+
+Reasoning Chains. Recognizing measurement system questions as multi-hop reasoning problems, we next investigate two strategies that push LLMs to reason: (1) Sequential Single Hops explicitly breaks down the problem for the LLM into individual reasoning steps and (2) Chain-of-Thought  $(CoT)$  prompts implicitly nudge the LLMs to reason by requiring an explanation for the answer.
+
+Sequential Single Hops (Seq). We first ask the LLMs to solve individual hops sequentially. To this end, we structure single-hop queries as illustrated in Figure 4. We feed the hops to the LLM in sequence, integrating the answer from the previous step into the next prompt: e.g., after obtaining the answer to "What is the exchange rate from USD to
+
+<table><tr><td colspan="4">Small Models</td><td colspan="6">Large Models</td></tr><tr><td colspan="4">Avg. Performance</td><td colspan="4">Avg. Performance</td><td colspan="2">Avg. Cost</td></tr><tr><td>Default Sys.</td><td colspan="3">Alternative Systems</td><td>Default Sys.</td><td colspan="3">Alternative Systems</td><td colspan="2">Alternative Sys.</td></tr><tr><td>NoReas.</td><td>ΔNoReas.</td><td>ΔSeq.</td><td>ΔCoT</td><td>NoReas.</td><td>ΔNoReas.</td><td>ΔSeq.</td><td>ΔCoT</td><td>ΔSeq.</td><td>ΔCoT</td></tr><tr><td colspan="10">Default: USD (Currency Systems)</td></tr><tr><td>53.43</td><td>-77.45%</td><td>-31.32%</td><td>-44.15%</td><td>74.67</td><td>-72.45%</td><td>-0.12%</td><td>+0.79%</td><td>+189%</td><td>+302%</td></tr><tr><td colspan="10">Default: Kilogram (Weight Systems)</td></tr><tr><td>32.43</td><td>-33.11%</td><td>-17.92%</td><td>-13.88%</td><td>54.33</td><td>-59.13%</td><td>-4.82%</td><td>-8.05%</td><td>+191%</td><td>+206%</td></tr><tr><td colspan="10">Default: Pound (Weight Systems)</td></tr><tr><td>59.08</td><td>-33.97%</td><td>-19.10%</td><td>-33.31%</td><td>75.61</td><td>-29.25%</td><td>-4.72%</td><td>-9.39%</td><td>+199%</td><td>+187%</td></tr><tr><td colspan="10">Default: Kilometer (Length Systems)</td></tr><tr><td>57.00</td><td>-63.64%</td><td>-31.82%</td><td>-41.74%</td><td>72.84</td><td>-57.93%</td><td>-7.88%</td><td>-3.98%</td><td>+180%</td><td>+210%</td></tr><tr><td colspan="10">Default: Mile (Length Systems)</td></tr><tr><td>68.13</td><td>-74.36%</td><td>-34.55%</td><td>-49.95%</td><td>92.46</td><td>-73.13%</td><td>-7.18%</td><td>-7.50%</td><td>+197%</td><td>+185%</td></tr></table>
+
+Table 4: Performance and Cost Increases across Reasoning Strategies: Performance: We report the percentage change relative to the default system (Default Sys.) without reasoning (NoReas., cf. §3.3). We highlight cells in red when the performance drop exceeds  $10\%$ . Cost Increase: We report the additional cost incurred compared to using no reasoning.
+
+VND in 2021?" we incorporate it into the subsequent query: "Convert <number> USD to VND."
+
+Chain-of-Thought (CoT). We explicitly prompt the LLM to detail its intermediate steps leading to the final answer. Specifically, we prompt the model with: "What is the GDP of Germany in ERN? Explain first. Be short."
+
+Evaluation. We report the change in MAPD relative to the default measurement system prompted without any reasoning (as discussed in Section 3.3), as shown in Table 4. Additionally, we examine the cost associated with these reasoning methods. For this purpose, we refer to the pricing provided by an LLM API provider<sup>3</sup> for the LLaMA 3.3 70B family and report the cost increase relative to the no-reasoning baseline (right part of Table 4). We further examine which measurement systems, in addition to the targeted system, are referenced in the CoT reasoning (see Figure 5).
+
+Results. Small LLMs experience a significant performance drop and show considerable instability, often failing to reason effectively, whereas Large models maintain stability across different measurement systems, with only slight performance deviations from the baseline. However, the cost of reasoning with Large models is considerably higher compared to the no-reasoning baseline. Sequential incur a cost up to  $199\%$  of the cost without
+
+reasoning, whereas CoT on average increases the test-time compute cost to  $300\%$  for currency conversions and around  $200\%$  for weight and length system conversions.
+
+Discussion. Prompting the model, explicitly (Seq) or implicitly (CoT) to reason does not benefit Small models. For Large models, reasoning does stabilize the performance to the level of the default measurement system without reasoning. This improvement comes at a significant cost: using reasoning increases test-time compute and, in turn, expenses by  $180\%$  to  $300\%$ , depending on the dimension of cultural context. While reasoning is an effective solution, its increased cost disproportionately affects users who rely on alternative measurement systems rather than the LLM's default. This discrepancy arises as models triangulate the conversion process step-by-step from the default measurement system. When faced with non-default measurement systems, LLMs engage in additional reasoning steps to convert the result from the default system (e.g., LLMs start from USD for currency conversion in  $91\%$  of the cases; see Figure 5), which demands additional computation. Consequently, users with cultural contexts that are misaligned with the LLMs' default (Anglo-centric context) face higher costs—this is yet another obstacle that disproportionately hinders access to state-of-the-art language technology for the least developed parts of the world.
+
+![](images/42c9e53f659a5386e48e0edfe1b39034ea57292478b9164ca62b2c6f361dd948.jpg)  
+Figure 5: Frequency of Measurement Systems as Starting Points in CoT Reasoning. We count how often each measurement system is chosen as the starting point for the model's CoT reasoning and report the proportion. We average this proportion across all models.
+
+# 5 Related Work
+
+Biases in Factual Retrieval. Recent research has examined LLMs' factual recall (Mallen et al., 2023; Kandpal et al., 2023; Sun et al., 2024; Tang et al., 2023, inter alia). While these models often provide correct answers, they can hallucinate (Huang et al., 2025). Studies also show that recall accuracy correlates with an entity's popularity—with higher error rates for lesser-known subjects (Mallen et al., 2023; Sun et al., 2024)—and reveal geographical biases along regional and income lines (Moayeri et al., 2024; Manvi et al., 2024).
+
+In contrast to prior work, our study shifts attention to the systems representing facts, noting that while measurement systems may influence depiction, the underlying facts remain unchanged. Park et al. (2022) demonstrate that small encoder models struggle with comparisons across different measurement systems. Additionally, Dinu et al. (2020) explicitly incorporates measurement systems into translation models. However, unlike these studies, our work focuses on decoder-based LLMs, emphasizing factual retrieval rather than simple comparisons or translations.
+
+Cultural Biases in LLMs. Kirk et al. (2024) underscore the disparities in how LLMs are used across diverse cultural contexts. However, numerous studies have quantified and identified cultural biases in LLMs and their multimodal extensions (Cao et al., 2024; Shi et al., 2024; Myung et al., 2024; Bui et al., 2024, inter alia). The consensus is that LLMs tend to align more closely with Western values, which limits their effectiveness in cross-cultural contexts. For instance, Lee et al. (2024) demonstrate that current LLMs achieve higher accuracies on hate speech labels derived from the Anglosphere. In our work, we further show that the same pattern models also favor measurement systems predominantly used in the Western world.
+
+Multi-Hop Reasoning. Research on multi-hop rea
+
+soning has produced numerous datasets and analyses (Yang et al., 2018; Fang et al., 2020; Chen et al., 2020; Trivedi et al., 2022, inter alia). Recent work has shifted toward implicit multi-hop questions that require models to recall and integrate facts without explicit reasoning steps (Press et al., 2023; Biran et al., 2024; Li et al., 2024; Yang et al., 2024). Although LLMs can connect internal knowledge, they still struggle when key reasoning details are absent (Press et al., 2023; Dziri et al., 2023). Moreover, while prior work mainly examines concatenated, independent facts, our analysis shows that natural multi-hop questions from measurement systems present unique challenges.
+
+# 6 Conclusion
+
+In this work, we show that LLMs exhibit significant variability in performance based on the measurement system embedded in the prompt for factual retrieval. Specifically, performance declines when the model encounters non-default measurement systems. We show that reasoning (e.g., CoT) improves the model's ability to handle alternative systems questions but it leads to an increase in test-time compute, which disproportionately affects users whose cultural contexts are misaligned with the LLMs' defaults.
+
+# Limitations
+
+In this study, we first prompt LLMs in English to examine how measurement systems affect factual retrieval accuracy. However, the interplay between language and measurement systems and its impact on performance remains an important avenue for future research. Users from underrepresented communities are much more likely to interact with dialogue systems in their native languages. Nonetheless, investigating how well LLMs generalize across measurement systems in English represents a 'best-case' scenario, given that these models perform best in English. Consequently, our core
+
+results would be even more pronounced in other languages, where LLMs are likely to mistranslate measurements more frequently and significantly, and incur higher inference costs due to tokenization into more subwords.
+
+Similarly, while our investigation examines a selected subset of measurement systems and units, we acknowledge the challenges of accessing comprehensive data—especially regarding systems still actively used within marginalized groups. Despite these constraints, our work successfully highlights key dynamics and interactions. We view this as a promising starting point that opens up opportunities for expanding the scope of research to include a broader array of measurement systems.
+
+Finally, we acknowledge that there are more effective tools for retrieving facts and performing unit conversions than solely relying on LLMs, such as RAG systems or dedicated conversion tools. However, we emphasize that users are more inclined to interact with a simple LLM chatbot for these seemingly straightforward tasks rather than resorting to more complex tools.
+
+# Ethical Statement
+
+The sources of our data, including the World Bank, IMF, and HDX, may reflect biases inherent in data collection methodologies, economic models, or reporting structures. These biases could influence the outcomes of analyses based on the dataset. Any model trained or analyzed solely on our dataset may inherit assumptions and limitations from the data sources.
+
+We use AI assistants, specifically o3-mini, to help edit sentences in our paper writing.
+
+# Acknowledgments
+
+The work of Minh Duc Bui and Katharina von der Wense was funded by the Carl Zeiss Foundation, grant number P2021-02-014 (TOPML project).
+
+# References
+
+Zeyuan Allen-Zhu and Yuanzhi Li. 2024. Physics of language models: Part 3.1, knowledge storage and extraction. Preprint, arXiv:2309.14316.  
+Eden Biran, Daniela Gottesman, Sohee Yang, Mor Geva, and Amir Globerson. 2024. Hopping too late: Exploring the limitations of large language models on multi-hop queries. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 14113-14130, Miami, Florida, USA. Association for Computational Linguistics.
+
+Minh Duc Bui, Katharina von der Wense, and Anne Lauscher. 2024. Multi3hate: Multimodal, multilingual, and multicultural hate speech detection with vision-language models. Preprint, arXiv:2411.03888.  
+Yong Cao, Yova Kementchedjhieva, Ruixiang Cui, Antonia Karamolegkou, Li Zhou, Megan Dare, Lucia Donatelli, and Daniel Hershcovich. 2024. Cultural adaptation of recipes. Transactions of the Association for Computational Linguistics, 12:80-99.  
+Wenhu Chen, Hanwen Zha, Zhiyu Chen, Wenhan Xiong, Hong Wang, and William Yang Wang. 2020. HybridQA: A dataset of multi-hop question answering over tabular and textual data. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1026-1036, Online. Association for Computational Linguistics.  
+John Dang, Shivalika Singh, Daniel D'souza, Arash Ahmadian, Alejandro Salamanca, Madeline Smith, Aidan Peppin, Sungjin Hong, Manoj Govindassamy, Terrence Zhao, Sandra Kublik, Meor Amer, Viraat Aryabumi, Jon Ander Campos, Yi-Chern Tan, Tom Kocmi, Florian Strub, Nathan Grinsztajn, Yannis Flet-berliac, and 26 others. 2024. Aya expanse: Combining research breakthroughs for a new multilingual frontier. Preprint, arXiv:2412.04261.  
+Georgiana Dinu, Prashant Mathur, Marcello Federico, Stanislas Lauly, and Yaser Al-Onaizan. 2020. Joint translation and unit conversion for end-to-end localization. In Proceedings of the 17th International Conference on Spoken Language Translation, pages 265-271, Online. Association for Computational Linguistics.  
+Nouha Dziri, Ximing Lu, Melanie Sclar, Xiang Lorraine Li, Liwei Jiang, Bill Yuchen Lin, Peter West, Chandra Bhagavatula, Ronan Le Bras, Jena D. Hwang, Soumya Sanyal, Sean Welleck, Xiang Ren, Allyson Ettinger, Zaid Harchaoui, and Yejin Choi. 2023. Faith and fate: limits of transformers on compositionality. In Proceedings of the 37th International Conference on Neural Information Processing Systems, NIPS '23, Red Hook, NY, USA. Curran Associates Inc.  
+Yuwei Fang, Siqi Sun, Zhe Gan, Rohit Pillai, Shuo-hang Wang, and Jingjing Liu. 2020. Hierarchical graph network for multi-hop question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 8823-8838, Online. Association for Computational Linguistics.  
+Neil Fantom and Umar Serajuddin. 2016. The world bank's classification of countries by income. Policy Research Working Paper 7528, World Bank. License: CC BY 3.0 IGO.  
+Mor Geva, Jasmijn Bastings, Katja Filippova, and Amir Globerson. 2023. Dissecting recall of factual associations in auto-regressive language models. In Proceedings of the 2023 Conference on Empirical Methods in
+
+Natural Language Processing, pages 12216-12235, Singapore. Association for Computational Linguistics.  
+Aaron Grattafori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad AlDahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, and 542 others. 2024. The llama 3 herd of models. Preprint, arXiv:2407.21783.  
+Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, and Ting Liu. 2025. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. ACM Trans. Inf. Syst., 43(2).  
+International Monetary Fund. 2024. International Financial Statistics: Exchange Rates. Accessed: 2024-02-13.  
+Nikhil Kandpal, Haikang Deng, Adam Roberts, Eric Wallace, and Colin Raffel. 2023. Large language models struggle to learn long-tail knowledge. In Proceedings of the 40th International Conference on Machine Learning, ICML'23. JMLR.org.  
+Hannah Rose Kirk, Alexander Whitefield, Paul Röttger, Andrew Bean, Katerina Margatina, Juan Ciro, Rafael Mosquera, Max Bartolo, Adina Williams, He He, Bertie Vidgen, and Scott A. Hale. 2024. The prism alignment dataset: What participatory, representative and individualised human feedback reveals about the subjective and multicultural alignment of large language models. In Advances in Neural Information Processing Systems, volume 37, pages 105236-105344. Curran Associates, Inc.  
+Paul R. Krugman. 1984. The international role of the dollar: Theory and prospect. In *Exchange Rate Theory and Practice*, pages 261-278. National Bureau of Economic Research (NBER).  
+Nayeon Lee, Chani Jung, Junho Myung, Jiho Jin, Jose Camacho-Collados, Juho Kim, and Alice Oh. 2024. Exploring cross-cultural differences in English hate speech annotations: From dataset construction to analysis. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 4205-4224, Mexico City, Mexico. Association for Computational Linguistics.  
+Zhaoyi Li, Gangwei Jiang, Hong Xie, Linqi Song, Defu Lian, and Ying Wei. 2024. Understanding and patching compositional reasoning in LLMs. In Findings of the Association for Computational Linguistics: ACL 2024, pages 9668-9688, Bangkok, Thailand. Association for Computational Linguistics.  
+Alex Mallen, Akari Asai, Victor Zhong, Rajarshi Das, Daniel Khashabi, and Hannaneh Hajishirzi. 2023.
+
+When not to trust language models: Investigating effectiveness of parametric and non-parametric memories. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 9802-9822, Toronto, Canada. Association for Computational Linguistics.  
+Rohin Manvi, Samar Khanna, Marshall Burke, David Lobell, and Stefano Ermon. 2024. Large language models are geographically biased. In Proceedings of the 41st International Conference on Machine Learning, ICML'24. JMLR.org.  
+Vaibhav Mavi, Anubhav Jangra, and Adam Jatowt. 2024. Multi-hop question answering. Preprint, arXiv:2204.09140.  
+Mazda Moayeri, Elham Tabassi, and Soheil Feizi. 2024. Worldbench: Quantifying geographic disparities in llm factual recall. In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency, FAccT '24, page 1211-1228, New York, NY, USA. Association for Computing Machinery.  
+Junho Myung, Nayeon Lee, Yi Zhou, Jiho Jin, Rifki Putri, Dimosthenis Antypas, Hsuvas Borkakoty, Eunsu Kim, Carla Perez-Almendros, Abinew Ali Ayele, Victor Gutierrez Basulto, Yazmin Ibanez-Garcia, Hwaran Lee, Shamsuddeen H Muhammad, Kiwoong Park, Anar Rzayev, Nina White, Seid Muhie Yimam, Mohammad Taher Pilehvar, and 3 others. 2024. Blend: A benchmark for llms on everyday knowledge in diverse cultures and languages. In Advances in Neural Information Processing Systems, volume 37, pages 78104-78146. Curran Associates, Inc.  
+Sungjin Park, Seungwoo Ryu, and Edward Choi. 2022. Do language models understand measurements? In *Findings of the Association for Computational Linguistics: EMNLP* 2022, pages 1782-1792, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.  
+Fabio Petroni, Tim Rocttäschel, Sebastian Riedel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, and Alexander Miller. 2019. Language models as knowledge bases? In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2463-2473, Hong Kong, China. Association for Computational Linguistics.  
+Ofir Press, Muru Zhang, Sewon Min, Ludwig Schmidt, Noah Smith, and Mike Lewis. 2023. Measuring and narrowing the compositionality gap in language models. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 5687-5711, Singapore. Association for Computational Linguistics.  
+Qwen, :: An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin
+
+Yang, Jiaxi Yang, Jingren Zhou, and 25 others. 2025. Qwen2.5 technical report. Preprint, arXiv:2412.15115.  
+Weiyan Shi, Ryan Li, Yutong Zhang, Caleb Ziems, Sunny Yu, Raya Horesh, Rogério Abreu De Paula, and Diyi Yang. 2024. CultureBank: An online community-driven knowledge base towards culturally aware language technologies. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 4996-5025, Miami, Florida, USA. Association for Computational Linguistics.  
+SimpleMaps. 2025. World cities database. https://simplemaps.com/data/world-cities. Accessed: 2025-05-28.  
+Kai Sun, Yifan Xu, Hanwen Zha, Yue Liu, and Xin Luna Dong. 2024. Head-to-tail: How knowledgeable are large language models (LLMs)? A.K.A. will LLMs replace knowledge graphs? In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 311-325, Mexico City, Mexico. Association for Computational Linguistics.  
+Liyan Tang, Tanya Goyal, Alex Fabbri, Philippe Laban, Jiacheng Xu, Semih Yavuz, Wojciech Kryscinski, Justin Rousseau, and Greg Durrett. 2023. Understanding factual errors in summarization: Errors, summarizers, datasets, error detectors. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 11626-11644, Toronto, Canada. Association for Computational Linguistics.  
+The World Bank. 2024. GDP per capita (current US). Accessed: 2024-02-13.  
+Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal. 2022. Musique: Multi-hop questions via single-hop question composition. Transactions of the Association for Computational Linguistics, 10:539-554.  
+U.S. Bureau of Labor Statistics. 2025. Consumer price index (cpi). Accessed: 2025-02-13.  
+Frank Wilcoxon. 1945. Individual comparisons by ranking methods. Biometrics Bulletin, 1(6):80-83.  
+World Food Program. 2025. Consumer price index (cpi). Accessed: 2025-02-13.  
+Sohee Yang, Elena Gribovskaya, Nora Kassner, Mor Geva, and Sebastian Riedel. 2024. Do large language models latently perform multi-hop reasoning? In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 10210-10229, Bangkok, Thailand. Association for Computational Linguistics.  
+Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William Cohen, Ruslan Salakhutdinov, and Christopher D. Manning. 2018. HotpotQA: A dataset for
+
+diverse, explainable multi-hop question answering. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2369-2380, Brussels, Belgium. Association for Computational Linguistics.
+
+# A Dataset
+
+# A.1 Measurement Systems
+
+The exhaustive list of measures used in the dataset are shown in Table 7.
+
+For ease of understanding, we additionally summarize the usage context and prevalence of each system in Table 5.
+
+# A.2 Terms of Use
+
+Our research is conducted in the public interest under GDPR, meeting the criteria for substantial academic research. To respect all data providers, our published work is licensed under CC BY-NC-ND 4.0. We use datasets with the following licenses:
+
+World Bank Group: The fiscal data is provided under an Access to Information policy and distributed as CC-BY 4.0.
+
+Exchange Rates: Data from the IMF is available for download, extraction, and use under its open terms.
+
+Food Prices: HDX and World Food Programme data use the Creative Commons Attribution for Intergovernmental Organisations license, while U.S. food prices from the Bureau of Labor Statistics are in the public domain.
+
+City Distances: Data from the Simplemaps domain is included in approximate form for reproducibility, with proper attribution as required by their terms.
+
+# B Experimental Setup
+
+# B.1 Prompts
+
+We report all prompts used in this paper.
+
+Base Prompts (No Reasoning). We create three base prompt variants for the no reasoning part:
+
+- Estimate the <MEASUREMENT> in year <YEAR>. Your answer must only include the <MEASUREMENT> as a decimal number without any abbreviations and explanation. Final number:  
+- Approximate the <MEASUREMENT> in year <YEAR>. Your answer must only include the <MEASUREMENT> as a decimal
+
+<table><tr><td>System</td><td>Usage Description</td><td>Usage Level</td></tr><tr><td colspan="3">Currencies</td></tr><tr><td>ALL</td><td>Currencies are used in every day life</td><td>High</td></tr><tr><td colspan="3">Weight System</td></tr><tr><td>Kilogram</td><td>Official SI base unit of mass</td><td>High</td></tr><tr><td>Pound</td><td>Imperial/US customary unit</td><td>High</td></tr><tr><td>Geun (Korea)</td><td>Korean traditional unit; still used in markets for meat.</td><td>Medium</td></tr><tr><td>Arrátel (Portuguese)</td><td>Historical Portuguese unit; obsolete.</td><td>Low</td></tr><tr><td>Funt (Russia)</td><td>Used historically in local trade and marketplaces; obsolete.</td><td>Low</td></tr><tr><td>Peittha</td><td>Burmese traditional unit, still used in Myanmar for local trade.</td><td>Medium</td></tr><tr><td>Troy pound</td><td>Specialized unit used for precious metals like gold and silver.</td><td>Medium</td></tr><tr><td>Jin (Mainland China)</td><td>Used in markets, and for measuring gold, silver and Chinese medicines.</td><td>Medium</td></tr><tr><td>Catty/Kati (Malaysian)</td><td>Traditional weight still used in some local markets in Southeast Asia.</td><td>Low</td></tr><tr><td>Sèr (Pakistani)</td><td>Traditional unit used historically in the Indian subcontinent; mostly obsolete.</td><td>Low</td></tr><tr><td colspan="3">Length Systems</td></tr><tr><td>Kilometer</td><td>Used in most countries as the official system.</td><td>High</td></tr><tr><td>Mile</td><td>Used primarily in the US for everyday measurements.</td><td>High</td></tr><tr><td>Li (China)</td><td>Traditional Chinese unit of length, historically used in land measurement.</td><td>Low</td></tr><tr><td>Ken (Japan)</td><td>Traditional Japanese unit of length, commonly used in architecture.</td><td>Medium</td></tr><tr><td>Eski Mil (Turkish)</td><td>Ottoman-era unit, obsolete in modern Turkey.</td><td>Low</td></tr><tr><td>Vyersta (Russian)</td><td>Historical Russian unit; obsolete.</td><td>Low</td></tr><tr><td>Cape Foot (South Africa)</td><td>Colonial-era unit, obsolete in contemporary use.</td><td>Low</td></tr><tr><td>Depa (Indonesian)</td><td>Traditional unit based on armspan, now rarely used.</td><td>Low</td></tr><tr><td>Argentine league (legua)</td><td>Historical unit used for long distances in rural Argentina; obsolete.</td><td>Low</td></tr><tr><td>Wa (Thai)</td><td>Traditional Thai unit still used for land measurement.</td><td>Medium</td></tr></table>
+
+Table 5: Overview of diverse measurement systems categorized by type (Currency, Weight, and Length), including their usage descriptions and estimated level of usage. We indicate High for systems that are officially used in the country, Medium for those that are not official but still used in specific local contexts, and Low for systems that are largely obsolete in everyday usage and are considered historical units.
+
+number without any abbreviations and explanation. Final number:
+
+- Guess the <MEASUREMENT> in the year <YEAR>. Your answer must only include the <MEASUREMENT> as a decimal number without any abbreviations and explanation. Final number:
+
+Base Prompts (CoT). To enable Chain-of- Thought (CoT) reasoning, we modify each base prompt. Specifically, we replace the instruction "Your answer must only include the <MEASUREMENT> as a decimal number without any abbreviations and explanation. Final number:" with "Explain first and then give your final answer with 'Answer: <Your answer in decimal number>''. Be short:"
+
+Variable Definitions. For different measurement systems, the variable  $< \text{MEASUREMENT}>$  is defined as follows:
+
+- Currency: <MEASUREMENT> = GDP per capita of <COUNTRY> in the currency <CURRENCY>
+
+- Weight Systems: <MEASUREMENT> = the price of <FOOD> per <WEIGHTSYSTEM> in <COUNTRY>
+
+- Length System: <MEASUREMENT> = distance between cities <CITY1> and <CITY2> in <COUNTRY> expressed in <LENGTH SYSTEM>
+
+# B.2 Model Details
+
+We employ the Hugging Face implementation for all models, which are run on two A100 80GB GPUs. Outputs are generated with a limit of 40 new tokens per request, and each large LLM processes a dataset in approximately 1.5 hours with a batch size of 64. For CoT prompting, we increase the token limit, which extends the processing time to nearly 24 hours.
+
+<table><tr><td></td><td>Llama 3.3</td><td colspan="2">Llama 3.1</td><td colspan="2">Qwen2.5</td><td colspan="2">Aya</td><td>Average</td></tr><tr><td></td><td>70B</td><td>70B</td><td>8B</td><td>72B</td><td>7B</td><td>32B</td><td>8B</td><td></td></tr><tr><td colspan="9">Currency (Fiscal Data)</td></tr><tr><td>Default = USD</td><td>71.51*</td><td>75.99</td><td>31.58*</td><td>76.51*</td><td>65.01*</td><td>63.08*</td><td>54.05*</td><td>62.53*</td></tr><tr><td>EUR</td><td>62.45</td><td>75.48</td><td>2.84</td><td>62.34</td><td>60.34</td><td>58.55</td><td>49.79</td><td>53.11</td></tr><tr><td>AUD</td><td>61.55</td><td>50.99</td><td>6.66</td><td>60.88</td><td>48.34</td><td>55.08</td><td>50.08</td><td>47.65</td></tr><tr><td>GBP</td><td>59.00</td><td>56.24</td><td>9.20</td><td>62.76</td><td>38.32</td><td>52.67</td><td>49.40</td><td>46.80</td></tr><tr><td>CAD</td><td>65.73</td><td>45.59</td><td>6.57</td><td>60.00</td><td>39.68</td><td>51.63</td><td>47.87</td><td>45.30</td></tr><tr><td>SGD</td><td>54.99</td><td>30.16</td><td>8.14</td><td>60.94</td><td>45.94</td><td>58.84</td><td>49.40</td><td>44.06</td></tr><tr><td>NZD</td><td>63.59</td><td>29.40</td><td>7.96</td><td>54.41</td><td>9.48</td><td>59.24</td><td>51.18</td><td>39.32</td></tr><tr><td>LOCAL</td><td>58.55</td><td>60.13</td><td>8.08</td><td>49.49</td><td>23.67</td><td>42.27</td><td>28.64</td><td>38.69</td></tr><tr><td>...</td><td>(all)</td><td>(all)</td><td>(all)</td><td>(all)</td><td>(all)</td><td>(all)</td><td>(all)</td><td>...</td></tr><tr><td>UZS</td><td>6.15</td><td>0.38</td><td>0.00</td><td>1.44</td><td>0.61</td><td>0.13</td><td>0.01</td><td>1.25</td></tr><tr><td>SDG</td><td>2.76</td><td>0.18</td><td>0.11</td><td>2.88</td><td>0.10</td><td>0.17</td><td>0.17</td><td>0.91</td></tr><tr><td colspan="9">Weight System (Food Prices)</td></tr><tr><td>Default=Kilogram</td><td>56.94*</td><td>52.83*</td><td>11.33</td><td>53.22*</td><td>42.67*</td><td>43.99*</td><td>31.74</td><td>41.82</td></tr><tr><td>Peitha</td><td>46.48</td><td>8.05</td><td>5.82</td><td>48.69</td><td>32.43</td><td>21.46</td><td>34.73</td><td>28.24</td></tr><tr><td>Sèr</td><td>41.28</td><td>17.49</td><td>10.85</td><td>26.31</td><td>30.10</td><td>24.44</td><td>28.81</td><td>25.61</td></tr><tr><td>Pound</td><td>29.04</td><td>28.44</td><td>14.68</td><td>28.81</td><td>27.56</td><td>26.43</td><td>23.40</td><td>25.48</td></tr><tr><td>Catty</td><td>23.95</td><td>10.90</td><td>7.57</td><td>33.71</td><td>32.27</td><td>26.40</td><td>23.13</td><td>22.56</td></tr><tr><td>Geun</td><td>26.52</td><td>11.91</td><td>10.04</td><td>24.36</td><td>31.05</td><td>25.54</td><td>24.58</td><td>22.00</td></tr><tr><td>Funt (Russian)</td><td>19.19</td><td>16.79</td><td>7.51</td><td>24.06</td><td>29.57</td><td>29.39</td><td>24.48</td><td>21.57</td></tr><tr><td>Jin</td><td>17.16</td><td>7.86</td><td>8.29</td><td>28.27</td><td>28.85</td><td>29.88</td><td>15.57</td><td>19.41</td></tr><tr><td>Arrátel</td><td>20.55</td><td>11.88</td><td>11.04</td><td>26.39</td><td>20.83</td><td>23.74</td><td>20.03</td><td>19.21</td></tr><tr><td>Troy Pound</td><td>6.21</td><td>7.61</td><td>15.56</td><td>7.58</td><td>18.97</td><td>18.77</td><td>17.09</td><td>13.11</td></tr><tr><td>Default=Pound</td><td>72.85</td><td>73.68</td><td>62.20*</td><td>80.29*</td><td>62.27*</td><td>66.61*</td><td>45.25</td><td>66.16</td></tr><tr><td>Funt (Russian)</td><td>68.15</td><td>69.61</td><td>55.22</td><td>74.82</td><td>56.25</td><td>61.25</td><td>46.61</td><td>61.70</td></tr><tr><td>Kilogram</td><td>75.32</td><td>74.74</td><td>45.12</td><td>60.82</td><td>54.66</td><td>42.11</td><td>44.56</td><td>56.76</td></tr><tr><td>Catty</td><td>68.54</td><td>70.46</td><td>34.57</td><td>66.56</td><td>49.91</td><td>47.11</td><td>36.49</td><td>53.38</td></tr><tr><td>Arrátel</td><td>51.63</td><td>63.14</td><td>38.99</td><td>71.05</td><td>59.68</td><td>48.24</td><td>31.09</td><td>51.98</td></tr><tr><td>Geun</td><td>51.59</td><td>62.01</td><td>33.56</td><td>55.71</td><td>46.59</td><td>32.37</td><td>36.24</td><td>45.44</td></tr><tr><td>Jin</td><td>49.62</td><td>63.81</td><td>22.85</td><td>55.12</td><td>33.75</td><td>38.81</td><td>25.99</td><td>41.42</td></tr><tr><td>Sèr</td><td>54.83</td><td>42.14</td><td>6.95</td><td>46.84</td><td>33.07</td><td>36.11</td><td>42.07</td><td>37.43</td></tr><tr><td>Troy Pound</td><td>10.70</td><td>30.20</td><td>57.83</td><td>25.12</td><td>58.28</td><td>32.14</td><td>42.14</td><td>36.63</td></tr><tr><td>Peitha</td><td>34.35</td><td>30.16</td><td>2.77</td><td>26.23</td><td>25.88</td><td>22.74</td><td>34.31</td><td>25.21</td></tr><tr><td colspan="9">Length System (City Distances)</td></tr><tr><td>Default=Kilometer</td><td>74.92*</td><td>74.85</td><td>51.95*</td><td>68.74*</td><td>59.26</td><td>66.66*</td><td>50.14*</td><td>63.79</td></tr><tr><td>Mile</td><td>70.81</td><td>72.42</td><td>36.25</td><td>64.39</td><td>56.25</td><td>58.14</td><td>26.27</td><td>54.93</td></tr><tr><td>Li</td><td>60.75</td><td>55.83</td><td>12.30</td><td>56.92</td><td>38.05</td><td>46.94</td><td>41.76</td><td>44.65</td></tr><tr><td>Vyersta</td><td>57.88</td><td>56.47</td><td>13.88</td><td>47.28</td><td>38.80</td><td>38.33</td><td>38.55</td><td>41.60</td></tr><tr><td>Eski Mil</td><td>4.78</td><td>7.64</td><td>0.11</td><td>10.07</td><td>35.81</td><td>22.41</td><td>15.24</td><td>13.72</td></tr><tr><td>Argentine League</td><td>8.58</td><td>2.55</td><td>0.20</td><td>11.71</td><td>35.71</td><td>8.60</td><td>0.41</td><td>9.68</td></tr><tr><td>Cape Foot</td><td>43.77</td><td>4.41</td><td>0.00</td><td>7.79</td><td>1.37</td><td>1.61</td><td>1.51</td><td>8.64</td></tr><tr><td>Wa</td><td>3.65</td><td>2.72</td><td>0.10</td><td>2.03</td><td>0.42</td><td>0.37</td><td>0.79</td><td>1.44</td></tr><tr><td>Depa</td><td>0.32</td><td>3.77</td><td>0.14</td><td>1.71</td><td>0.23</td><td>0.50</td><td>0.99</td><td>1.10</td></tr><tr><td>Ken</td><td>0.27</td><td>0.70</td><td>0.07</td><td>0.60</td><td>4.72</td><td>0.27</td><td>0.67</td><td>1.04</td></tr><tr><td>Default=Millioner</td><td>94.54*</td><td>92.72</td><td>80.96</td><td>90.11*</td><td>58.40</td><td>77.25</td><td>55.92</td><td>78.56</td></tr><tr><td>Kilometer</td><td>91.35</td><td>90.31</td><td>78.40</td><td>83.88</td><td>65.10*</td><td>77.24</td><td>57.84</td><td>77.73</td></tr><tr><td>Vyersta</td><td>75.83</td><td>80.50</td><td>10.43</td><td>73.97</td><td>45.05</td><td>53.45</td><td>53.66</td><td>56.13</td></tr><tr><td>Li</td><td>61.57</td><td>56.29</td><td>27.66</td><td>55.53</td><td>41.82</td><td>47.46</td><td>40.84</td><td>47.31</td></tr><tr><td>Eski Mil</td><td>4.78</td><td>7.12</td><td>11.84</td><td>33.70</td><td>62.27</td><td>21.59</td><td>14.74</td><td>22.29</td></tr><tr><td>Cape Foot</td><td>43.72</td><td>25.95</td><td>0.01</td><td>0.84</td><td>0.12</td><td>0.39</td><td>0.10</td><td>10.16</td></tr><tr><td>Argentine League</td><td>4.24</td><td>2.89</td><td>0.34</td><td>1.87</td><td>47.48</td><td>2.02</td><td>0.61</td><td>8.49</td></tr><tr><td>Ken</td><td>0.27</td><td>12.56</td><td>0.03</td><td>0.26</td><td>0.15</td><td>0.76</td><td>0.75</td><td>2.11</td></tr><tr><td>Wa</td><td>5.15</td><td>5.48</td><td>0.19</td><td>0.52</td><td>0.17</td><td>0.31</td><td>0.27</td><td>1.73</td></tr><tr><td>Depa</td><td>0.22</td><td>5.46</td><td>0.14</td><td>0.26</td><td>0.12</td><td>0.40</td><td>0.38</td><td>1.00</td></tr></table>
+
+Table 6: Disparity and Performance of LLMs Across Different Measurement Systems: The first row (shaded in gray) within each group represents the model's default measurement system. We report the performance in terms of MAPD (cf. §3.3). The best performance in each column (higher is better) are shown in bold. Disparity: The best values are additionally marked with *, if statistically significant differences  $(p < 0.001)$  compared to the second best system exists. Cells highlighted in red indicate cases where a significant difference was observed.
+
+Currency (Fiscal Data)  
+
+<table><tr><td>Currency code</td><td>Description</td><td>Conversion (to USD)</td><td>Currency code</td><td>Description</td><td>Conversion (to USD)</td></tr><tr><td>AED</td><td>United Arab Emirates Dirham</td><td>3.6725</td><td>KMF</td><td>Comorian Franc</td><td>418.0677243</td></tr><tr><td>AFN</td><td>Afghan Afghani</td><td>77.41418667</td><td>KRW</td><td>South Korean Won</td><td>1149.108333</td></tr><tr><td>ALL</td><td>Albanian Lek</td><td>104.0547222</td><td>KWD</td><td>Kuwaiti Dinar</td><td>0.301641667</td></tr><tr><td>AMD</td><td>Armenian Dram</td><td>503.6105556</td><td>KZT</td><td>Kazakhstani Tenge</td><td>425.9513889</td></tr><tr><td>ARS</td><td>Argentine Peso</td><td>95.18694444</td><td>LBP</td><td>Lebanese Pound</td><td>1507.5</td></tr><tr><td>AUD</td><td>Australian Dollar</td><td>1.337372732</td><td>LKR</td><td>Sri Lankan Rupee</td><td>201.3890667</td></tr><tr><td>AZN</td><td>Azerbaijani Manat</td><td>1.7</td><td>LYD</td><td>Libyan Dinar</td><td>4.523094444</td></tr><tr><td>BAM</td><td>Bosnia-Herzegovina Convertible</td><td>1.6673125</td><td>MDL</td><td>Moldovan Leu</td><td>17.66990278</td></tr><tr><td></td><td>Mark</td><td></td><td>MAD</td><td>Moroccan Dirham</td><td>9.045236111</td></tr><tr><td>BDT</td><td>Bangladeshi Taka</td><td>85.24412794</td><td>MGA</td><td>Malagasy Ariary</td><td>3854.072778</td></tr><tr><td>BGN</td><td>Bulgarian Lev</td><td>1.661877778</td><td>MKD</td><td>Macedonian Denar</td><td>52.55698889</td></tr><tr><td>BHD</td><td>Bahraini Dinar</td><td>0.376</td><td>MMK</td><td>Myanmar Kyat</td><td>1610.540741</td></tr><tr><td>BIF</td><td>Burundian Franc</td><td>1978.428889</td><td>MOP</td><td>Macaoese Pataca</td><td>8.008919444</td></tr><tr><td>BND</td><td>Brunei Dollar</td><td>1.345794444</td><td>MUR</td><td>Mauritian Rupee</td><td>41.72418333</td></tr><tr><td>BOB</td><td>Bolivian Boliviano</td><td>6.91</td><td>MXN</td><td>Mexican Peso</td><td>20.654375</td></tr><tr><td>BRL</td><td>Brazilian Real</td><td>5.368375</td><td>MYR</td><td>Malaysian Ringgit</td><td>4.143145833</td></tr><tr><td>BWP</td><td>Botswana Pula</td><td>11.19945833</td><td>MZN</td><td>Mozambican Metical</td><td>66.0225</td></tr><tr><td>BYN</td><td>Belarusian Ruble</td><td>2.533175</td><td>NGN</td><td>Nigerian Naira</td><td>406.2341667</td></tr><tr><td>BZD</td><td>Belize Dollar</td><td>2</td><td>NIO</td><td>Nicaraguaan Cordoba</td><td>35.20105833</td></tr><tr><td>CAD</td><td>Canadian Dollar</td><td>1.251083333</td><td>NOK</td><td>Norwegian Krone</td><td>8.631944444</td></tr><tr><td>CDF</td><td>Congolese Franc</td><td>1989.948356</td><td>NPR</td><td>Nepalese Rupee</td><td>118.7686111</td></tr><tr><td>CHF</td><td>Swiss Franc</td><td>0.9150375</td><td>NZD</td><td>New Zealand Dollar</td><td>1.414869444</td></tr><tr><td>CLP</td><td>Chilean Peso</td><td>779.4822222</td><td>OMR</td><td>Omani Rial</td><td>0.3845</td></tr><tr><td>CNY</td><td>Chinese Yuan</td><td>6.457772222</td><td>PAB</td><td>Panamanian Balboa</td><td>1</td></tr><tr><td>COP</td><td>Colombian Peso</td><td>3784.358056</td><td>PEN</td><td>Peruvian Nuevo Sol</td><td>3.892861111</td></tr><tr><td>CRC</td><td>Costa Rican Colon</td><td>625.7229167</td><td>PHP</td><td>Philippine Peso</td><td>49.43672222</td></tr><tr><td>CVE</td><td>Cape Verdean Escudo</td><td>93.765085</td><td>PKR</td><td>Pakistan Rupee</td><td>164.1562869</td></tr><tr><td>CZK</td><td>Czech Republic Koruna</td><td>21.67958333</td><td>PLN</td><td>Polish Zloty</td><td>3.889841667</td></tr><tr><td>DJF</td><td>Djiboutian Franc</td><td>177.721</td><td>PYG</td><td>Paraguayan Guarani</td><td>6676.519722</td></tr><tr><td>DKK</td><td>Danish Krone</td><td>6.317175</td><td>QAR</td><td>Qatari Rial</td><td>3.64</td></tr><tr><td>DOP</td><td>Dominican Peso</td><td>57.25499444</td><td>RON</td><td>Romanian Leu</td><td>4.192530556</td></tr><tr><td>DZD</td><td>Algerian Dinar</td><td>135.6026</td><td>RSD</td><td>Serbian Dinar</td><td>99.94511944</td></tr><tr><td>EGP</td><td>Egyptian Pound</td><td>15.65245278</td><td>RUB</td><td>Russian Ruble</td><td>73.53540833</td></tr><tr><td>ETB</td><td>Ethiopian Birr</td><td>44.25473611</td><td>RWF</td><td>Rwandan Franc</td><td>991.6896025</td></tr><tr><td>EUR</td><td>Euro</td><td>0.84978685</td><td>SAR</td><td>Saudi Riyal</td><td>3.75</td></tr><tr><td>GBP</td><td>British Pound Sterling</td><td>0.72827027</td><td>SDG</td><td>Sudanese Pound</td><td>299.4778056</td></tr><tr><td>GEL</td><td>Georgian Lari</td><td>3.25225</td><td>SEK</td><td>Swedish Krona</td><td>8.649225</td></tr><tr><td>GHS</td><td>Ghanaian Cedi</td><td>5.851411111</td><td>SGD</td><td>Singapore Dollar</td><td>1.345016667</td></tr><tr><td>GNF</td><td>Guinean Franc</td><td>9933.2195</td><td>SYP</td><td>Syrian Pound</td><td>1256</td></tr><tr><td>GTQ</td><td>Guatemala Quetzal</td><td>7.7412875</td><td>THB</td><td>Thai Baht</td><td>32.01021944</td></tr><tr><td>HKD</td><td>Hong Kong Dollar</td><td>7.77677778</td><td>TND</td><td>Tunisian Dinar</td><td>2.805069444</td></tr><tr><td>HNL</td><td>Honduran Lempira</td><td>24.0325</td><td>TRY</td><td>Turkish Lira</td><td>9.808511111</td></tr><tr><td>HUF</td><td>Hungarian Forint</td><td>304.9938889</td><td>TTD</td><td>Trinidad and Tobago Dollar</td><td>6.747011111</td></tr><tr><td>IDR</td><td>Indonesian Rupiah</td><td>14333.63958</td><td>TWD</td><td>New Taiwan Dollar</td><td>28.05294444</td></tr><tr><td>ILS</td><td>Israeli New Sheqel</td><td>3.224777778</td><td>TZS</td><td>Tanzanian Shilling</td><td>2296.881064</td></tr><tr><td>INR</td><td>Indian Rupee</td><td>73.82430833</td><td>UAH</td><td>Ukrainian Hryvnia</td><td>27.27136389</td></tr><tr><td>IQR</td><td>Iraqi Dinar</td><td>1450</td><td>UGX</td><td>Ugandan Shilling</td><td>3601.599952</td></tr><tr><td>IRR</td><td>Iranian Rial</td><td>42000</td><td>USD</td><td>US Dollar</td><td>1</td></tr><tr><td>ISK</td><td>Icelandic Krona</td><td>125.9425</td><td>UYU</td><td>Uruguayan Peso</td><td>43.51508333</td></tr><tr><td>JMD</td><td>Jamaican Dollar</td><td>150.3607333</td><td>UZS</td><td>Uzbekistan Som</td><td>10646.02833</td></tr><tr><td>JOD</td><td>Jordanian Dinar</td><td>0.71</td><td>VND</td><td>Vietnamese Dong</td><td>23177.22222</td></tr><tr><td>JPY</td><td>Japanese Yen</td><td>109.6466667</td><td>XAF</td><td>CFA Franc BEAC</td><td>557.4236325</td></tr><tr><td>KES</td><td>Kenyan Shilling</td><td>110.2038889</td><td>XOF</td><td>CFA Franc BCEAO</td><td>557.4236325</td></tr><tr><td>KHR</td><td>Cambodian Riel</td><td>4097.125</td><td>ZAR</td><td>South African Rand</td><td>14.955069</td></tr></table>
+
+Weight System  
+
+<table><tr><td>Measure</td><td>Description</td><td>Conversion (in KG)</td><td>Measure</td><td>Description</td><td>Conversion (in KG)</td></tr><tr><td>Kilogram (kg)</td><td></td><td>1</td><td>Geun (Korea)</td><td></td><td>0.6</td></tr><tr><td>Pound (lb)</td><td></td><td>0.453592</td><td>Troy pound</td><td></td><td>0.37324</td></tr><tr><td>Arrátel (Portugese)</td><td></td><td>0.459</td><td>Funt (Russia)</td><td></td><td>0.409517</td></tr><tr><td>catty/kati (Malaysian)</td><td></td><td>0.60479</td><td>Sèr (80 Tôlia, Pakistani)</td><td></td><td>0.93310</td></tr><tr><td>Jin (Mainland China)</td><td></td><td>0.5</td><td>Peitha (Burmamese)</td><td></td><td>1.63293</td></tr></table>
+
+Length System  
+
+<table><tr><td>Measure</td><td>Description</td><td>Conversion (in KM)</td><td>Measure</td><td>Description</td><td>Conversion (in KM)</td></tr><tr><td>Kilometer (km)</td><td></td><td>1</td><td>Mile (mi)</td><td></td><td>1.60934</td></tr><tr><td>Li (China)</td><td></td><td>0.5</td><td>Eski Mil (Turkish)</td><td></td><td>1.89435</td></tr><tr><td>Vyersta (Russian)</td><td></td><td>1.0668</td><td>Ken (Japan)</td><td></td><td>0.001818</td></tr><tr><td>Cape Foot (South Africa)</td><td></td><td>0.0003148584</td><td>Depa (Indonesian)</td><td></td><td>0.0016</td></tr><tr><td>Argentine league (legua)</td><td></td><td>5.572</td><td>Wa (Thai)</td><td></td><td>0.002</td></tr></table>
+
+Table 7: Measurement System Description. Description of measures used in the dataset.

ACL/2025/On Generalization across Measurement Systems_ LLMs Entail More Test-Time Compute for Underrepresented Cultures/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5ddb5166e1fc3c3ecc8da602aa6d93b9b03f069599424b60e2132a7e2fe23afb
+size 1337710

ACL/2025/On Generalization across Measurement Systems_ LLMs Entail More Test-Time Compute for Underrepresented Cultures/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5ee401be52b365c101ff0e4a2beb63e9d44399f71b36dde614f281ebadd4c7f6
+size 412765

ACL/2025/On Many-Shot In-Context Learning for Long-Context Evaluation/ede27d5b-5aa5-4444-838b-ee0bfc3b5d5f_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bbfdc9ac72defb52fac967137761ef3d65bd9fb35abce3357886dc4097358618
+size 183620

ACL/2025/On Many-Shot In-Context Learning for Long-Context Evaluation/ede27d5b-5aa5-4444-838b-ee0bfc3b5d5f_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f1c14116939cae48a58b53faebe81d040baadbad93ac4b405eb1edb7bc3181ec
+size 206303

ACL/2025/On Many-Shot In-Context Learning for Long-Context Evaluation/ede27d5b-5aa5-4444-838b-ee0bfc3b5d5f_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4af57bf6f561d3a751d0d19442c961add280b0473b35ea69b796bf4a0890c993
+size 2855383

ACL/2025/On Many-Shot In-Context Learning for Long-Context Evaluation/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:df48ed35e114722f6786254e601d9388952cf7584202fbe48e47d56cad98921a
+size 3139873

ACL/2025/On Many-Shot In-Context Learning for Long-Context Evaluation/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e289dbfc0e78d34f658f077dcc49e5dde5cea2c0f10807ea8164acd629525f3a
+size 809697

ACL/2025/On Support Samples of Next Word Prediction/d8ab6303-8390-447e-b23b-0b88e1def99d_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:46208b65d3ee42fcf7098b4a011c8fd1f98f3f7d8771fa9e683a4e5c5223746e
+size 91162

ACL/2025/On Support Samples of Next Word Prediction/d8ab6303-8390-447e-b23b-0b88e1def99d_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ca17566ed24a911496eafa4a894f7f325ae20e632cb8c3a5c42993ed2b6a5203
+size 106656

ACL/2025/On Support Samples of Next Word Prediction/d8ab6303-8390-447e-b23b-0b88e1def99d_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c39815d27db18b9da29a110392c256ce4c03a05e9a7eb5b75d636ef984e884ec
+size 1397534

ACL/2025/On Support Samples of Next Word Prediction/full.md

@@ -0,0 +1,390 @@
+# On Support Samples of Next Word Prediction
+
+Yuqian Li $^{1*}$ , Yupei Du $^{2*}$ , Yufang Liu $^{1}$
+
+Feifei Feng $^{3}$ , Mou Xiao Feng $^{3}$ , Yuanbin Wu $^{1}$ ,
+
+<sup>1</sup>School of Computer Science and Technology, East China Normal University,
+
+2Utrecht University, the Netherlands,
+
+3Midea Group,
+
+liyuqian44@gmail.com, y.du@uu.nl, ybwu@cs.ecnu.edu.cn
+
+# Abstract
+
+Language models excel in various tasks by making complex decisions, yet understanding the rationale behind these decisions remains a challenge. This paper investigates data-centric interpretability in language models, focusing on the next-word prediction task. Using representer theorem, we identify two types of support samples—those that either promote or deter specific predictions. Our findings reveal that being a support sample is an intrinsic property, predictable even before training begins. Additionally, while non-support samples are less influential in direct predictions, they play a critical role in preventing overfitting and shaping generalization and representation learning. Notably, the importance of non-support samples increases in deeper layers, suggesting their significant role in intermediate representation formation. These insights shed light on the interplay between data and model decisions, offering a new dimension to understanding language model behavior and interpretability.
+
+# 1 Introduction
+
+Language models make decisions. From selecting an answer from QA benchmarks to generating reasoning steps for grade school math problems in GSM8k, language models are applauded and criticized for their decision-making ability. Like all other AI systems, understanding how decisions are made is an important research topic.
+
+Support of decisions can be explored in different parts of language models. One common attribution is the patterns of neuron activations (mechanistic interpretability): some neurons are more sensitive to a specific decision and some are not. The circuit formed by active neurons mechanistically explains
+
+the model's decision (Elhage et al., 2021). Mechanistic interpretability targets on model parameters (and hidden states), however, it is not the only possible place to explore the rationale behind decisions. From the perspective of compression, parameters are the compressed version of training data. So, a natural question is whether we can trace model decisions back to their remote origin, the data.
+
+We follow this point of view and study datacentric interpretability of language models. Concretely, we will focus on decisions about choosing the next word (the essential decision of language models), and try to answer the question "Which samples contribute the most when the model decides to predict a token  $v$ ?" For general machine learning problems, there are two typical methods: one is based on counterfactual argument (a sample is important if the decision will change if it is removed from the training set (Koh and Liang, 2017)), another is based on representation theorems (a sample is important if it occupies a large part of prediction parameters (Yeh et al., 2018)). The first one requires computing Hessians of model parameters which could be expensive for analyzing large language models (Grosse et al., 2023b). We thus adopt the more efficient representation-based methods. Specifically,
+
+- We illustrate the contribution of different samples to model parameters based on a simple representation theorem for the next word predictor. The results remind us of two types of support samples (those contribute a lot), one attracts the predictor to predict  $v$ , while another pushes the predictor away from predicting  $v$ .  
+- We show that, given a dataset and a model configuration, being a support sample is an intrinsic property of that sample, in the sense that 1) non-support samples have limited influence on learning support samples, and 2)
+
+it is possible to predict (with  $80\%$  accuracy) whether a sample is a support one at the very early phase of language model training, even without any training.
+
+- We find that though non-support samples are less noticeable in the next-word predictor, it may play an important role in controlling generalization and representations learning. Specifically, we observe that 1) without any non-support samples, learning the predictor will suffer from unnecessary overfitting, and a small number of non-support samples can greatly alleviate the problem. 2) the proportion of non-support samples increases as the layer goes deeper, which means some of them are support samples of (thus contributing a lot to) intermediate representations.
+
+# 2 Next Word Predictor Representation
+
+We aim to find important samples to support a language model's next-word decisions. To measure importance, we first connect parameters (in particular, prediction heads at the last layer) and training samples with a representation theorem, then support samples are defined to be those samples contributing significantly to parameters.
+
+Denote  $(\mathbf{x},\mathbf{y})$  to be a training sample, where  $\mathbf{x} = x_{1},x_{2},\ldots ,x_{t - 1}$  is a sentence prefix,  $\mathbf{y} = x_{t}\in V$  is the target token to predict given the prefix, and  $V$  is the vocabulary.  $D = \{(\mathbf{x}_i,\mathbf{y}_i)\}_{i = 1}^N$  is a set of samples. We consider a decomposition of language models with a representation function  $\phi (\mathbf{x})\in \mathbb{R}^{d}$  and a token prediction function  $f(\mathbf{x}) = \arg \max_vp(v|\mathbf{x})$  ,where  $p(v|\mathbf{x}) =$ $\frac{1}{Z}\exp \theta_v^T\phi (\mathbf{x})$  .Let  $\theta_{\phi},\theta_{v}$  be the parameter of  $\phi (\cdot)$  and  $f(\cdot),\theta_V = [\theta_{v_1};\theta_{v_2};\dots ;\theta_{v|V|}]$  pack all predictor parameters (also called language model heads), and  $\theta = [\theta_V;\theta_\phi ]$  pack all parameters.
+
+Theorem 1 (Crammer and Singer 2001; Yeh et al. 2018). Assume  $\theta$  is a stationary point of the loss function  $L(\theta, D) = -\frac{1}{N} \sum_{i=1}^{N} \log p(\mathbf{y}_i | \mathbf{x}_i) + \lambda \| \theta \|_2^2$ , then  $\theta_v$  equals
+
+$$
+\theta_ {v} = \frac {1}{2 N \lambda} \sum_ {i = 1} ^ {N} (\mathbb {1} (\mathbf {y} _ {i} = v) - p (v | \mathbf {x} _ {i})) \phi (\mathbf {x} _ {i}) \tag {1}
+$$
+
+where  $\mathbb{1}(\cdot)$  equals 1 when the argument is true, 0 otherwise.
+
+The theorem gives a decomposition of the predictor  $\theta_v$  using training samples  $\phi(\mathbf{x})$ . The coefficients of samples  $\alpha_i \triangleq \mathbb{1}(\mathbf{y}_i = v) - p(v|\mathbf{x}_i)$
+
+describe their importance in  $\theta_v$ . A large  $\alpha_i$  implies a bigger influence, and it also implies the sample is difficult to learn: a large  $\alpha_i$  means  $p(v|\mathbf{x}_i)$  is small when  $\mathbf{y}_i = v$ , or  $p(v|\mathbf{x}_i)$  is large when  $\mathbf{y}_i \neq v$ , in both cases the probability of the ground truth token is small.
+
+Definition 2. The support samples of predicting token  $v$  are defined to be  $S_{v} = \{(\mathbf{x}_{i},\mathbf{y}_{i})|\tau \leq |\alpha_{i}|\}$ . The support samples of the full language model are  $S = \cup_{v}S_{v}$ . The non-support samples are  $\bar{S}$ .<sup>2</sup>
+
+We have seen that support samples are roughly hard-to-predict samples, here we can also inspect the opposite side by considering the relationship between non-support samples and easy-to-predict samples. Specifically, we consider  $M = \{(\mathbf{x},\mathbf{y})|f(\mathbf{x}) = \mathbf{y},p(v|\mathbf{x})\geq \gamma \}$ , which are samples correctly predicted after training, and defined to be "memorized" by language models (Tirumala et al., 2022). The following fact shows that the memorized samples are a kind of non-support samples.
+
+Claim 3. If  $\gamma, \tau \geq 0.5$ , the non-support samples  $\bar{S}$  contains  $M$ .
+
+Proof. When  $\gamma \geq 0.5$ ,  $p(\mathbf{y}|\mathbf{x}) \geq \tau$  implies  $f(\mathbf{x}) = \mathbf{y}$ , and  $M = \{(\mathbf{x}, \mathbf{y}) | p(v|\mathbf{x}) \geq \gamma\}$ .
+
+At the same time, for any  $(\mathbf{x},\mathbf{y})\in M$  , if there exists a  $v$  such that  $(\mathbf{x},\mathbf{y})\in S_v$
+
+- if  $v = \mathbf{y}$ ,  $\alpha = 1 - p(\mathbf{y}|\mathbf{x}) < 0.5 \leq \tau$ ,  
+- if  $v \neq \mathbf{y}$ , since  $p(\mathbf{y}|\mathbf{x}) > 0.5$ ,  $\alpha = p(v|\mathbf{x}) < 0.5 \leq \tau$ .
+
+There are contradictions  $\alpha <  \tau$  in both cases, thus for all  $v$ $(\mathbf{x},\mathbf{y})\notin S_{v}$
+
+# 3 Support Samples at First Glance
+
+We now profile support samples. Our exemplar language model is trained from scratch with GPT-2 architecture (117M parameters, with  $l = 12$  Transformer layers and hidden size  $d = 768$ ) and wikitext-2 dataset (2.37M training samples). The vocabulary size  $|V| = 50257$ . We will investigate models with more parameters (345M, 774M, 1.5B) and larger training set (wikitext-103 with 117M
+
+![](images/016c6645b627fe9f1d5251846fc5f0fc6065f888da9c8fed044c3ea4e046cbe6.jpg)  
+Figure 1: Number distribution of support samples for each token  $v$ . Blue bars are the top  $42\%$  tokens with the most support samples. Orange bars are the remaining  $58\%$  tokens which have less than 10 support samples. The table shows the number of the top 5 tokens.
+
+samples) in specific experiments, and report results on larger models in Appendix E.
+
+Number of support samples. Our first observation is that, with sufficient training (loss converges) and standard model selection (best validation set performances), among 2.37M samples, there are 1.29M support samples  $(54\%)$ . The proportion is remarkably high: more than half of the samples are important for the next-word predictor. The larger amount of support samples implies fewer patterns are discovered during training: the predictor fails to make the correct decision by keeping a few representative samples.
+
+By depicting the number of support samples for individual  $v \in V$  (Figure 1), we further find the distribution is highly screwed: very few  $v$  contribute a large number of support samples. It implies that though the overall support samples are many, prediction patterns of most tokens are clear (i.e.,  $58\%$  tokens have support samples less than 10). We additionally provide a concrete example of the support and non-support samples for a specific  $v$  in Appendix C.
+
+POS tags of support samples. To inspect support samples more closely, we roughly annotate POS tag of the next token y given its prefix x. 5,
+
+![](images/855366213ad29a8cc0d972fd6cb5ec47e469beaf0412b17eadc46fca4e95e7f9.jpg)  
+Figure 2: Proportion of support samples in different POS tags. "Random" means that a random sampling on the full dataset will pick  $54\%$  (which is the proportion of support samples) support samples. The figure shows that the proportions of support samples in two POS-tag groups significantly differ from the random proportion.
+
+and see whether samples with different target POS tags tokens will have different probabilities to be support samples. Figure 2 shows that  $86\%$  of verbs are support samples while the proportion is only  $20\%$  for punctuations. In general, support samples are more prevalent in semantically rich POS tags (e.g., ADV, NUM, ADJ, VERB, NOUN), while less common in semantically light tags (e.g., PART, AUX, DET, ADP, PUNCT). It indicates that relatively, the predictor keeps more information about verbs than punctuations in its parameters.
+
+Two types of support samples By revisiting the representation of  $\theta_v$ , we can group support samples into two types. Specifically, Equation 1 can be rewrite as (dropping  $(N\lambda)^{-1}$ ),
+
+$$
+\sum_ {\mathbf {y} _ {i} = v} (1 - p (v | \mathbf {x} _ {i})) \phi (\mathbf {x} _ {i}) + \sum_ {\mathbf {y} _ {i} \neq v} (- p (v | \mathbf {x} _ {i})) \phi (\mathbf {x} _ {i}) \tag {2}
+$$
+
+For each target token  $v$ , we define support samples from the first summation to be Type-1, and the second summation Type-2. Namely, Type-1 are support samples with the same target token as  $v$ , but their prediction confidence is low. Type-2 are samples with different target tokens, but they are confident in predicting the wrong target  $v$ .
+
+To understand the roles of Type-1 and Type-2 samples, we conduct the following counterfactual argument: what if we remove (subtract) one support sample type of  $v$  from all prediction parameters  $\theta_V$ . Let  $(\mathbf{x}',\mathbf{y}')$  be a testing sample. Empiri
+
+<table><tr><td rowspan="2">Removed Set</td><td colspan="2">v=&quot;hens&quot;</td><td colspan="2">v=&quot;ction&quot;</td></tr><tr><td>full-set</td><td>y=v</td><td>full-set</td><td>y=v</td></tr><tr><td>-∅</td><td>3.28</td><td>0.24</td><td>3.28</td><td>0.56</td></tr><tr><td>-Sv</td><td>4.45</td><td>2.37</td><td>3.52</td><td>1.50</td></tr><tr><td>-Type-1 of Sv</td><td>3.35</td><td>16.73</td><td>3.28</td><td>33.09</td></tr><tr><td>-Type-2 of Sv</td><td>4.75</td><td>0.00</td><td>3.90</td><td>0.00</td></tr><tr><td>-random</td><td>3.75</td><td>0.27</td><td>3.80</td><td>0.51</td></tr></table>
+
+Table 1: Removing support samples of  $v =$  "hens" and "ction". The number of support samples  $\left| {S}_{v}\right|  = {10}$  ,and half of them are Type-1. When removing the samples, we subtract them from all  ${\theta }_{v}$  . We consider four sets to remove,  ${S}_{v}$  ,Type-1/Type-2 of  ${S}_{v}$  ,and a random subset with the same number. We report the loss of all training samples (full-set) and loss of a subset which contains samples with  $v$  as the target token  $\left( {\mathbf{y} = v}\right)$  .
+
+cally, we assume  $\phi (\mathbf{x}_i)^T\phi (\mathbf{x}^\prime)\geq 0$  for all  $i$
+
+If we subtract Type-1 samples of  $v$ , the logit  $\theta_v^T\phi (\mathbf{x}')$  will be smaller since a positive term  $\sum_{\mathbf{y} = v}(1 - p(\mathbf{y}_i|\mathbf{x}_i))\phi (\mathbf{x}_i)^T\phi (\mathbf{x}')$  is dropped. With a softmax activation, the probability of predicting  $v$  could be greatly decreased, and the model resists predicting  $v$ . Similarly, when subtracting Type-2 samples, a negative term is dropped, and the model tends to predict  $v$ . Therefore, we would think that the existence of Type-1 support samples pull the predictor towards predicting  $v$ , while Type-2 samples push the predictor away from predicting  $v$ .
+
+Table 1 lists a concrete example for which we conduct the above subtracting experiments. The results show that, for these two when removing Type-1, predictions on  $v$  fail, and predictions on other samples can be kept (even improved, compared with random subtracting: 3.35 vs. 3.75, and 3.28 vs. 3.80). When removing Type-2, predictions on  $v$  are perfect, while other samples are negatively affected. It shows that holding Type-2 support samples is crucial for the predictor, which is new to the perplexity-based data selection principle (only Type-1 samples are important).
+
+To further understand Type-2, we draw a network to illustrate connections of next-token targets. In Figure 3, we link tokens according to Type-2 supporting relations: each directed edge from  $v$  to  $u$  indicates that a sample with  $v$  as the target token is the Type-2 support sample of  $u$ . We also list hubs with most in-degrees and out-degrees.
+
+![](images/3ad4471f4ac2c345a22ffbef5f0e9459d84cdc21ccfb57c58538446e2224edb2.jpg)  
+Figure 3: An illustration of token relations according to Type-2. The number of nodes is 1690 (3% of  $V$ , not all tokens have Type-2 support samples). The number of edges is 1993. We lists tokens with top  $\mathrm{k}(=5)$  out-degree and in-degree in the network.
+
+<table><tr><td>vout-degree</td><td>,30</td><td>@28</td><td>and27</td><td>.22</td><td>in17</td></tr><tr><td>vin-degree</td><td>=242</td><td>of153</td><td>s144</td><td>,134</td><td>@84</td></tr></table>
+
+# 4 Importance of Support Samples
+
+Support samples contribute the most to  $\theta_v$  (LM heads). This leads us to pose a question: If the support samples are defined to be the most important samples, can we only use them (ignore those non-support samples) during training?
+
+We examine alternative ways of removing non-support samples: removing with sampling 1) hard: removing all non-support samples; 2) random; uniformly removing the same number of samples as the non-support; 3) soft: using the coefficient  $\alpha_{i}$  as the probability for sampling, meaning that some non-support samples will be retained. We experiment with two training configurations: only training LM heads (other parameters inherited from the original trained LM) and full model training. The result is shown in Table 2.
+
+For training LM heads, the answer is: Yes, we can remove non-support samples (around  $40\% - 50\%$  of the samples) while keeping the same performances, with weighted sampling or even random sampling. However, for training full models, the answer is No.
+
+An interesting phenomenon is that soft sampling outperforms random sampling for training LM heads (fixed representation space), but the opposite occurs when training full model. The former tells non-support samples are not necessary for
+
+learning heads (in the sense of soft removing). The latter implies non-support samples are important for learning representations (parts of the model beyond the heads). Otherwise, the model trained only with support samples will perform like only head tuning with support samples. A deeper exploration of this point is presented in Section 6.
+
+Curiously, for tuning heads, hard sampling performs worse than soft sampling.(more or even exclusively support samples does not lead to better performance, and meanwhile random sampling with fewer support samples performs worse.) Observing the changes in support samples, we find that although hard sampling slightly reduces original support samples (1.29→1.18), it turns more unseen non-support samples into support(1.29→1.90). This makes the predictor's hyperplane unclearer (more support samples), and leads to a performance drop. It means without any non-support samples, learning the predictor will suffer from unnecessary overfitting to support. A small number of non-support samples can greatly alleviate the problem (soft better than hard), while support samples are still key for tuning heads (soft better than random). Regarding full model training, the lack of non-support samples poses a more significant issue, thus random sampling surpasses soft sampling, because non-support samples are more important for parts of the model beyond the heads as mentioned in the previous paragraph.
+
+See Table 2 again, a pattern seems to emerge: the fewer the support samples, the lower the loss. We already see that the learned next-word predictor has many support samples. The next question is: Is it possible to make the hyperplane clearer by reducing some support samples?
+
+We try more data, various regularization methods (L2 and dropout). Figure 4 first shows that the trends of loss and support proportion are largely consistent. And more data can help reducing support samples to make the hyperplane clearer and the predictor better, while L2 and embedding dropout fail.
+
+In conclusion, we get 1) For tuning heads, support samples are key, but some non-support samples are also necessary. We can remove non-support samples with soft sampling. 2) Non-support samples are important for learning representations. 3) Fewer support samples indicate the clearer hyperplane and better performance of the predictor. More data vitals.
+
+# 5 Predicting Support Samples
+
+We have another discovery when training with subsets (not shown in Table 2): the loss on original support samples are almost the same for models trained with and without non-support samples (6.17 vs. 6.43).<sup>7</sup>
+
+The result implies that non-support samples have limited influence on better fitting support samples, and it roughly suggests that the hardness of support samples might be "intrinsic". To make the problem clear, we cast an interesting question, can we predict whether a sample is a support sample or not without training the language model?
+
+Surprisingly, we have a positive answer. We will show that given initial parameters, with simple classifiers (linear and MLP) and a limited number of training samples (annotations of support and non-support samples), one could recognize samples with high accuracy. It may confirm our "intrinsic" property intuition. The following are details.
+
+Denote  $\zeta (\mathbf{x},\mathbf{y},\theta^{(t)})$  to be a feature map which extracts features of a sample  $(\mathbf{x},\mathbf{y})$  according to model parameter  $\theta^{(t)}$  at checkpoint  $t$  (especially, we are interested in the initial state  $\theta^{(0)})$ .  $h_\eta (\cdot)$  is a binary classifier (with parameter  $\eta$ ) which takes a feature vector  $\mathbf{z} = \zeta (\mathbf{x},\mathbf{y},\theta^{(t)})$  and outputs  $\mathbf{s}\in$ $\{0,1\}$  indicating whether the input is a support sample or not. To learn the classifier, we need a training set  $E = \{(\mathbf{z}_j,\mathbf{s}_j)\}_{j = 1}^K$ . We experiment with the following configurations.
+
+The feature map  $\zeta$  We consider three types of features at checkpoint  $t$ ,
+
+- the last hidden vectors  $\phi_{\theta^{(t)}}(\mathbf{x}) \in \mathbb{R}^d$ , which is  $\phi(\mathbf{x})$  evaluated with checkpoint  $\theta^{(t)}$ . Recalling that support samples are defined according to  $p(v|\mathbf{x})$  which is a generalized linear model of  $\phi(\mathbf{x})$ , last hidden vectors seem to be a natural and effective choice.  
+- concatenation of all hidden vectors  $[\phi_{\theta(t)}^1(\mathbf{x}); \phi_{\theta(t)}^2(\mathbf{x}); \ldots; \phi_{\theta(t)}^l(\mathbf{x})] \in \mathbb{R}^{dl}$ , where  $\phi_{\theta(t)}^j(\mathbf{x})$  is the hidden vector of layer  $j$ . The feature includes intermediate representations.  
+- gradient features  $\nabla_{\theta}p(\mathbf{y}|\mathbf{x})|_{\theta^{(t)}}$ . Instead of directly using the gradients, which are in the
+
+<table><tr><td rowspan="2">Method</td><td rowspan="2">Test Loss</td><td colspan="2"># Support Samples(M)</td></tr><tr><td>new training set (after removal)</td><td>original training set</td></tr><tr><td>-</td><td>5.08</td><td>1.29</td><td>1.29</td></tr><tr><td>hard</td><td>5.64</td><td>1.29→1.18</td><td>1.29→1.90</td></tr><tr><td>soft</td><td>5.13</td><td>0.87→0.87</td><td>1.29→1.33</td></tr><tr><td>random</td><td>5.18</td><td>0.70→0.74</td><td>1.29→1.38</td></tr></table>
+
+(a) only training LM heads  
+
+<table><tr><td rowspan="2">Method</td><td rowspan="2">Test Loss</td><td colspan="2"># Support Samples(M)</td></tr><tr><td>new training set (after removal)</td><td>original training set</td></tr><tr><td>-</td><td>5.08</td><td>1.29</td><td>1.29</td></tr><tr><td>hard</td><td>6.57</td><td>1.29→1.28</td><td>1.29→2.36</td></tr><tr><td>soft</td><td>5.69</td><td>0.87→1.01</td><td>1.29→1.76</td></tr><tr><td>random</td><td>5.47</td><td>0.70→0.73</td><td>1.29→1.45</td></tr></table>
+
+(b) full model training
+
+Table 2: Performance of the predictor and Change of support samples when removing non-support samples from original training set with various sampling methods. -: no removing (original training set). hard: removing all non-support samples; random: randomly removing the same number of samples as the non-support; soft: using the coefficient  $\alpha_{i}$  as the probability for weighted sampling, meaning that some non-support samples will be retained. We count support samples on two types of dataset: one is new training set (after removal) as the initial number of support samples varies for each method, another is original training set.
+
+![](images/f9ef00e13b008ea1aa115fe72639e6ee18a39f083aa20d9b9c67768695371011.jpg)  
+Figure 4: The change of Loss and Support proportion on different training sets. Data Size: we sample subsets of different sizes from wikitext-103.L2: we set the weight decay parameter to adjust L2 regularization during training. Dropout: we try different dropout rates at the embedding layer.
+
+![](images/96e152fe4492adbed8bde3d2e8e8efe965ea2f320f7ea03efb5a4a9995efc7bd.jpg)
+
+![](images/b8865d242cc0905ee5192b507af513e0d223c5e111709e537ab11a61f9d47ac3.jpg)
+
+extremely high dimension space, we perform a random projection of the gradient vector to a low dimension (4096).
+
+# Classifier We consider two classifiers,
+
+- linear classifiers,  $h_{\eta}(\mathbf{z}) = \eta^{T}\zeta (\mathbf{x},\mathbf{y},\theta^{(t)})$  The learnable parameter  $\eta$  equals the dimension of input features.  
+- MLPs,  $h_{\eta}(\mathbf{z}) = w^{T}W_{2}\sigma (W_{1}\zeta (\mathbf{x},\mathbf{y},\theta^{(t)}))$ , where  $\sigma$  is a non-linear activation function, and  $\eta = [w;W_1;W_2]$ . We try different model capacities by varying the size of  $W_{1},W_{2}$ .
+
+Training set We collect training set  $E$  for this classification task from the training samples of the language model  $D$ . We first train the language
+
+model with  $D$ , and get annotations of support/non-support samples according to the last checkpoint. We split the set to training, validation and testing set with 8:1:1 ratio. We also test smaller versions of the training set.
+
+Checkpoints We extract features from different checkpoints of the language model. Two special checkpoints are,
+
+- checkpoint  $\theta^{(0)}$  which is the initial language model without any training.  
+- the last checkpoint  $\theta^{(T)} = \theta$  which is the parameter defining the gold standard annotation of support/non-support samples. We may think the performances of classifiers at the last checkpoint are performances upper bounds: a classifier can see all necessary information for recognizing support samples.
+
+As  $\theta^{(0)}$  could be hard, we are also interested in the performances at the early checkpoints of training.
+
+The experimental results (Figure 5) show that with MLP and gradient feature, the prediction accuracy can reach  $80\%$  without training, just  $5\%$  shy of the upper limit (last checkpoint). The classifier of the last checkpoint still cannot achieve  $100\%$  accuracy, suggesting that the cut between support and non-support samples is not clear. If we allow extracting features from early training phases, the prediction accuracy can be better.
+
+Performance improves with more data and more parameters (MLP with increasing intermediate dimension) but eventually levels off. The upper limit of gradient feature is around  $80\%$ , of last hidden state is about  $65\%$ , and of all hidden states is  $66\%$ . Gradient outperforms hidden states, particularly in the early training stages. Even with more parameters, hidden states still cannot catch up.
+
+In brief, we successfully predict whether a sample is a support sample without training (80% accuracy) only using a simple classifier (MLP) with a randomly projected gradient feature.
+
+# 6 Non-support Samples
+
+In Section 2, from the counterfactual (removing non-support samples), we preliminarily infer that non-support samples are crucial for learning model parameters beyond the heads. Now we further show that non-support samples play a role in learning representations.
+
+In Transformers architectures, higher-layer representations are typically better than lower-layer ones. Therefore, if higher-layer representations "contain" more non-support samples, could this indicate that non-support samples contribute to learning representations? We probe each layer of LM (learning new heads for each hidden states with the original training object of the LM).
+
+Figure 6 describes the number and POS distribution of non-support samples across different layers.10 Overall, representations of the LM "contain" more non-support samples at higher layers. In the first six layers, the number of non-support samples remains relatively low and stable, with a sudden surge observed from the fifth to the sixth layer. The growth in higher layers is markedly more pronounced than in lower layers. For POS, the LM
+
+shows some recognition of NUM in the first layer. Notably, PUNCT is heavily contained at the sixth layer.
+
+Now we can answer: non-support samples indeed play a role in learning representations. Conversely, they can also be viewed as a visualization of representational capacity: the LM's representational ability undergoes a qualitative leap starting at the sixth (half) layer.
+
+![](images/6b41c7c96b158e26896582a6837ca5d46e05762e87f230adb48eb4fa817eea64.jpg)  
+Figure 6: The number and POS distribution of non-support samples across different layers of the LM. Overall, representations of the LM "contain" more non-support samples at higher layers. In the first six layers, the number of non-support samples remains relatively low and stable, with a sudden surge observed from the fifth to the sixth layer. The growth in higher layers is markedly more pronounced than in lower layers. For POS, the LM shows some recognition of NUM in the first layer. Notably, PUNCT is heavily contained at the sixth layer.
+
+# 7 Related Work
+
+Difficult instances and model training Many studies have investigated the relationship between the difficulty of training instances and their role for model's learning, and observe that keeping harder instances and a small amount of easy instances can maintain or even improve model's generalization performance. A wide range of difficulty measures have been adopted, such as "forgetting" behavior, i.e., after which this training example is always correctly predicted (Toneva et al., 2019), training loss magnitude (Jiang et al., 2021; Swayamdipta et al., 2020), gradient norm (Paul et al., 2021), the number of layers needed for prediction (Baldock et al., 2021), perplexity (Kwok et al., 2024), and hidden representations' distance to other instances (Sorscher et al., 2022). Different from
+
+![](images/d78dcc59e0ca5f476429d9d9d46344a57704ea37ad7d57c5af7b9869daea2bce.jpg)  
+Figure 5: Performance of the classifier under various variables. Left: Performance changes across different training stages. The accuracy achieves nearly  $80\%$  even without training, and only improves to  $85\%$  at the last/best step. Middle: Performance improves with more parameters (MLP with increasing intermediate dimension) but eventually levels off. The upper limit of gradient is around  $80\%$ , of last hidden state is about  $65\%$ , and of all hidden states is  $66\%$ . Due to the varying feature dimensions (last hidden state: 768; all hidden states: 9984; gradient: projected to 4096) and significant differences in parameter counts, only the portion within 350k parameters is shown in the figure for clarity. In practice, all hidden states can achieve  $66\%$  with more parameters. Right: Performance enhances with the growth of training data. All three figures demonstrate that, in terms of features, gradient outperforms hidden states, particularly in the early training stages. Even with more parameters, hidden states still cannot catch up.
+
+![](images/d5b67d777e3e3616e4e161c66c2e0f8368565f66c0a8d17ae5c36c90b465c74e.jpg)
+
+![](images/8ebb19c54b2eb186c1484279187c5ecdc653c6f92d62de6291b8c6194804bc29.jpg)
+
+these works that aim to identify small but informative subsets for training, we study the specific roles of support and non-support examples in learning representations and decision boundaries.
+
+Predictive Data Attribution Predictive data attribution seeks to answer the question: what would have happened if we had trained on a different dataset? (Madry et al., 2024). Most existing approaches rely on the gradients of individual instances.11 One line of research uses influence functions (Koh and Liang, 2017) and its approximations (Koh et al., 2019; Park et al., 2023; Grosse et al., 2023a). However, other studies have observed them to be fragile (Basu et al., 2021) and do not answer the question of leave-one-out retraining (Bae et al., 2022). The other line of works approximates the training dynamics, by unrolling the updates during each step (Hara et al., 2019; Pruthi et al., 2020; Bae et al., 2024).
+
+# 8 Conclusion
+
+This work has taken a data-centric approach, to study the training samples that contribute the most to the prediction of each token using a representation theorem. Concretely, we make three contributions,
+
+- We observe that support samples constitute over half of the training set. Interestingly, some predictions are supported by a large
+
+number of samples, while others are supported by only a few (§3). suggesting language models may employ two distinct modes of prediction. Furthermore, we identify two types of support samples: those that help predict the correct tokens and those that prevent incorrect token predictions.
+
+- We further study the roles of support samples in training using removing-retraining experiments. We find that support and non-support samples are respectively more useful for learning prediction heads and backbone representations ( $\S 4$  and  $\S 6$ ). Interestingly, we observe that a few non-support samples are always needed for successful training.  
+- Inspired by the observation that non-support samples do not contribute to learning support samples, we investigate whether the difficulty of support samples is intrinsic by predicting them without training language models. Surprisingly, we observe that using only the gradients of a randomly initialized language model can predict support samples with over  $80\%$  accuracy, validating our hypothesis (§5).
+
+# 9 Limitation
+
+On the ambiguity between support and non-support samples. In section 6, we have noted that even when using features from the last checkpoint for training, the classifier's accuracy on the test set is still not perfect. This suggests that the
+
+boundary between support and non-support samples is not well-defined. Directly dividing training samples into support and non-support based on a threshold may have inherent issues.
+
+Beyond argmax: rethinking decoding and sample importance. The argmax decoding enters the representation theorem through the MLE loss function(Theorem 1), which is common in LM training. If the language model is trained in a different object function, both the representation of next-word prediction heads and the definition of important samples could be different.
+
+However, for inference, the analysis is not limited to the argmax decoding. For example, when estimating importance scores, one could use sampling instead of argmax prediction. The sampling-based decoding may better reflect the nature of language data, which often does not align with a deterministic majority-vote interpretation. This may help address the concern that the observed next token  $y$  should not be assumed to be the argmax outcome of the underlying distribution.
+
+In terms of the sparsity of next-word distribution and the role of argmax decoding. We would think that the learning of language model tries to establish an approximate mapping between context-dependent semantics of a token  $v$  (i.e., presentations of various prefixes ended with  $v$ ) and context-independent semantics of the token (i.e., presentations of  $v$  in vocabulary). We would also think that the next-word prediction task is more like a nearest-neighbor search in the learned vocabulary space.
+
+# 10 Acknowledgement
+
+The authors wish to thank all reviewers for their helpful comments and suggestions. The corresponding authors are Yuanbin Wu and Yufang Liu.
+
+# References
+
+Juhan Bae, Wu Lin, Jonathan Lorraine, and Roger Baker Grosse. 2024. Training data attribution via approximate unrolling. In The Thirty-eighth Annual Conference on Neural Information Processing Systems.  
+Juhan Bae, Nathan Hoyen Ng, Alston Lo, Marzyeh Ghassemi, and Roger Baker Grosse. 2022. If influence functions are the answer, then what is the question? In Advances in Neural Information Processing Systems.  
+Robert John Nicholas Baldock, Hartmut Maennel, and Behnam Neyshabur. 2021. Deep learning through
+
+the lens of example difficulty. In Advances in Neural Information Processing Systems.  
+Samyadeep Basu, Phil Pope, and Soheil Feizi. 2021. Influence functions in deep learning are fragile. In International Conference on Learning Representations.  
+Koby Crammer and Yoram Singer. 2001. On the algorithmic implementation of multiclass kernel-based vector machines. J. Mach. Learn. Res., 2:265-292.  
+Nelson Elhage, Neel Nanda, Catherine Olsson, Tom Henighan, Nicholas Joseph, Ben Mann, Amanda Askell, Yuntao Bai, Anna Chen, Tom Conerly, Nova DasSarma, Dawn Drain, Deep Ganguli, Zac Hatfield-Dodds, Danny Hernandez, Andy Jones, Jackson Kernion, Liane Lovitt, Kamal Ndousse, Dario Amodei, Tom Brown, Jack Clark, Jared Kaplan, Sam McCandlish, and Chris Olah. 2021. A mathematical framework for transformer circuits. Transformer Circuits Thread. https://transformercircuits.pub/2021/framework/index.html.  
+Roger Grosse, Juhan Bae, Cem Anil, Nelson Elhage, Alex Tamkin, Amirhossein Tajdini, Benoit Steiner, Dustin Li, Esin Durmus, Ethan Perez, Evan Hubinger, Kamile Lukosiiute, Karina Nguyen, Nicholas Joseph, Sam McCandlish, Jared Kaplan, and Samuel R. Bowman. 2023a. Studying large language model generalization with influence functions. Preprint, arXiv:2308.03296.  
+Roger B. Grosse, Juhan Bae, Cem Anil, Nelson Elhage, Alex Tamkin, Amirhossein Tajdini, Benoit Steiner, Dustin Li, Esin Durmus, Ethan Perez, Evan Hubinger, Kamile Lukosiute, Karina Nguyen, Nicholas Joseph, Sam McCandlish, Jared Kaplan, and Samuel R. Bowman. 2023b. Studying large language model generalization with influence functions. CoRR, abs/2308.03296.  
+Satoshi Hara, Atsushi Nitanda, and Takanori Maehara. 2019. Data cleansing for models trained with sgd. In Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc.  
+Andrew Ilyas, Sung Min Park, Logan Engstrom, Guillaume Leclerc, and Aleksander Madry. 2022. Data-models: Predicting predictions from training data. In ICML.  
+Ziheng Jiang, Chiyuan Zhang, Kunal Talwar, and Michael C Mozer. 2021. Characterizing structural regularities of labeled data in overparameterized models. In Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 5034-5044. PMLR.  
+Pang Wei Koh and Percy Liang. 2017. Understanding black-box predictions via influence functions. In Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pages 1885-1894. PMLR.
+
+Pang Wei W Koh, Kai-Siang Ang, Hubert Teo, and Percy S Liang. 2019. On the accuracy of influence functions for measuring group effects. In Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc.
+
+Devin Kwok, Nikhil Anand, Jonathan Frankle, Gintare Karolina Dziugaite, and David Rolnick. 2024. Dataset difficulty and the role of inductive bias. arXiv preprint arXiv:2401.01867.
+
+Aleksander Madry, Andrew Ilyas, Logan Engstrom, Sung Min (Sam) Park, and Kristian Georgiev. 2024. Data attribution at scale. Tutorial presented at the 41st International Conference on Machine Learning (ICML).
+
+Sung Min Park, Kristian Georgiev, Andrew Ilyas, Guillaume Leclerc, and Aleksander Madry. 2023. Trak: Attributing model behavior at scale. In International Conference on Machine Learning (ICML).
+
+Mansheej Paul, Surya Ganguli, and Gintare Karolina Dziugaite. 2021. Deep learning on a data diet: Finding important examples early in training. In Advances in Neural Information Processing Systems.
+
+Garima Pruthi, Frederick Liu, Satyen Kale, and Mukund Sundararajan. 2020. Estimating training data influence by tracing gradient descent. In Advances in Neural Information Processing Systems, volume 33, pages 19920-19930. Curran Associates, Inc.
+
+Ben Sorscher, Robert Geirhos, Shashank Shekhar, Surya Ganguli, and Ari S. Morcos. 2022. Beyond neural scaling laws: beating power law scaling via data pruning. In Advances in Neural Information Processing Systems.
+
+Swabha Swayamdipta, Roy Schwartz, Nicholas Lourie, Yizhong Wang, Hannaneh Hajishirzi, Noah A. Smith, and Yejin Choi. 2020. Dataset cartography: Mapping and diagnosing datasets with training dynamics. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 9275-9293, Online. Association for Computational Linguistics.
+
+Kushal Tirumala, Aram H. Markosyan, Luke Zettlemoyer, and Armen Aghajanyan. 2022. Memorization without overfitting: Analyzing the training dynamics of large language models. In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9, 2022.
+
+Mariya Toneva, Alessandro Sordoni, Remi Tachet des Combes, Adam Trischler, Yoshua Bengio, and Geoffrey J. Gordon. 2019. An empirical study of example forgetting during deep neural network learning. In International Conference on Learning Representations.
+
+Chih-Kuan Yeh, Joon Sik Kim, Ian En-Hsu Yen, and Pradeep Ravikumar. 2018. Representer point selection for explaining deep neural networks. In Advances in Neural Information Processing Systems 31: Annual Conference on Neural Information Processing Systems 2018, NeurIPS 2018, December 3-8, 2018, Montreal, Canada, pages 9311-9321.
+
+# A Proof of Theorem 1
+
+Proof. We first compute derivatives of  $\log p(\mathbf{y}|\mathbf{x})$
+
+$$
+\frac {\mathrm {d} \log p (\mathbf {y} | \mathbf {x})}{\mathrm {d} \theta_ {v}} = \frac {\mathrm {d} \left(\theta_ {\mathbf {y}} ^ {T} \phi (\mathbf {x}) - \log Z\right)}{\mathrm {d} \theta_ {v}}.
+$$
+
+For the normalizer  $Z = \sum_{v'} \exp \left(\theta_{v'}^T \phi(\mathbf{x})\right)$ , we have,
+
+$$
+\frac {\mathrm {d} \log Z}{\mathrm {d} \theta_ {v}} = \frac {1}{Z} \exp \left(\theta_ {v} ^ {T} \phi (\mathbf {x})\right) \cdot \phi (\mathbf {x}) = p (v | \mathbf {x}) \cdot \phi (\mathbf {x}).
+$$
+
+Therefore,
+
+$$
+\begin{array}{l} \frac {\mathrm {d} \log p (\mathbf {y} | \mathbf {x})}{\mathrm {d} \theta_ {v}} = \left\{ \begin{array}{c c} (1 - p (\mathbf {y} | \mathbf {x}))   \phi (\mathbf {x}), & \mathbf {y} = v \\ - p (\mathbf {y} | \mathbf {x}) \phi (\mathbf {x}), & \mathbf {y} \neq v \end{array} \right. \\ = \left(\mathbb {1} (\mathbf {y} = v) - p (\mathbf {y} | \mathbf {x})\right) \phi (\mathbf {x}). \\ \end{array}
+$$
+
+Next, since  $\theta$  is a stationary point,  $\frac{\partial L(\theta, D)}{\partial \theta_v} = 0$ , we have
+
+$$
+\begin{array}{l} - \frac {1}{N} \sum_ {i = 1} ^ {N} (\mathbb {1} (\mathbf {y} _ {i} = v) - p (v | \mathbf {x} _ {i})) \phi (\mathbf {x} _ {i}) + 2 \lambda \theta_ {v} = 0 \\ \Rightarrow \theta_ {v} = \frac {1}{2 N \lambda} \sum_ {i = 1} ^ {N} (\mathbb {1} (\mathbf {y} _ {i} = v) - p (v | \mathbf {x} _ {i})) \phi (\mathbf {x} _ {i}) \\ \end{array}
+$$
+
+![](images/a35d7e409869548f4b4eaed01d3badcf183d7d6bb8a6ec3cb0880cce868086f2.jpg)
+
+# B Rationale Behind the Choice of Hyperparameter  $\tau$
+
+We set  $\tau = 0.9$  empirically, based on the following rationale:
+
+Distribution of sample importance scores  $\alpha$ . As shown in Figure 7, we observe a drastic increase of  $\alpha$  from  $9.27\%$  to  $54.16\%$  from  $\tau = 0.8$  to  $\tau = 0.9$ , suggesting a natural threshold for different regions.
+
+Ablation studies for threshold parameter  $\tau$  sensitivity. We report the number of support samples with different  $\tau$  as illustrated in Figure 8. The sharp decrease of the number of support samples from 0.9 to 1.0 further justifies our choice.
+
+![](images/94ac5641f1a431f9da3c47173763ab989bcc132d111fad02b3316ae93b2ba873.jpg)  
+Figure 7: Distribution of sample importance scores  $\alpha$ .
+
+![](images/0ecbbd8df90afdc5278789e743825953aaa54bbdf7525aa7c2d49e4fc11bb83e.jpg)  
+Figure 8: The number of support samples with different  $\tau$ .
+
+Softmax sharpen probabilities. Since probabilities are processed through softmax, they tend to be quite sharp. Actually, using a temperature parameter can make the model less sensitive to threshold choices, resulting in a smoother distribution.
+
+Actually, the choice of threshold does not affect the analysis framework in the work: all methods can be generalized to any threshold.
+
+On the other hand, threshold selection is practically significant. From a practitioner's viewpoint, with a fixed budget, determining an appropriately sized support set and exploring how to adjust thresholds for different support set sizes are valuable directions for future work.
+
+# C Concrete Example
+
+We present some support and non-support samples of etts, where the target token is etts itself, in Table 3. In our framework, etts is typically a suffix ob
+
+![](images/974d14c2768e6112e15599ad2239eb87dcc4b58c8ec767b573ca0d4a94e43d95.jpg)  
+Figure 9: The number and POS distribution of memorized samples across different layers.
+
+served in proper nouns such as Burnetts, Corbetts, and Plunketts. Our statistics reveal that Burnetts and Corbetts(of support samples) appear once each in the training data, while Plunketts(of non-support samples) appears 68 times. This example suggests that less frequent patterns in the training data are more likely to become support to the LM heads' parameters.
+
+<table><tr><td>Type</td><td>Sample</td></tr><tr><td rowspan="2">Support</td><td>... The northern edge of the Plunketts Creek drainage basin is formed by Burnetts</td></tr><tr><td>... They are mainly composed of granite that has weathered into more rounded hills with many long scree slopes on their flanks. The highest point of these hills is Glamaig, one of only two Corbetts</td></tr><tr><td>Non-support</td><td>... Much of the Plunketts Creek valley is composed of various glacial deposits, chiefly alluvium. Although the Plunketts</td></tr></table>
+
+Table 3: Partial support and non-support samples of the token etts.
+
+# D Layer-wise Non-support/Memorized Samples Distribution
+
+Figure 9 presents the probing result of memorized samples, which reveals a trend consistent with non-support samples as shown in Figure 6.
+
+# E Scaling Up
+
+We scale our experiments to 345M(GPT-2-Medium architecture), 774M(GPT-2-Large architecture), and 1.5B(GPT-2-XL architecture) models. Our findings remain consistent across these scales as follows.
+
+<table><tr><td rowspan="2">Method</td><td rowspan="2">Test Loss</td><td colspan="2"># Support Samples(M)</td></tr><tr><td>new training set (after removal)</td><td>original training set</td></tr><tr><td>-</td><td>5.02</td><td>1.13</td><td>1.13</td></tr><tr><td>hard</td><td>5.62</td><td>1.13→0.98</td><td>1.13→1.77</td></tr><tr><td>soft</td><td>5.07</td><td>0.73→0.70</td><td>1.13→1.25</td></tr><tr><td>random</td><td>5.12</td><td>0.57→0.53</td><td>1.13→1.23</td></tr></table>
+
+(a) only training LM heads  
+
+<table><tr><td rowspan="2">Method</td><td rowspan="2">Test Loss</td><td colspan="2"># Support Samples(M)</td></tr><tr><td>new training set (after removal)</td><td>original training set</td></tr><tr><td>-</td><td>5.02</td><td>1.13</td><td>1.13</td></tr><tr><td>hard</td><td>6.67</td><td>1.13→1.12</td><td>1.13→2.36</td></tr><tr><td>soft</td><td>5.77</td><td>0.87→0.70</td><td>1.13→1.78</td></tr><tr><td>random</td><td>5.51</td><td>0.65→0.53</td><td>1.13→1.51</td></tr></table>
+
+(b) full model training
+
+Table 4: Performance of the predictor and Change of support samples of the 774M model when removing non-support samples from original training set with various sampling methods.
+
+![](images/1ee88ffa456b7bf520a69462a8c8b22a0f77c8282536afae5fed09414d079765.jpg)  
+Figure 10: Distribution of samples importance scores  $\alpha$  across models of various sizes trained on the wikitext-2 dataset.
+
+Distribution of sample importance scores  $\alpha$ . We compare the distribution across models of various sizes trained on the wikitext-2 dataset and find them similar, highlighting the scalability of our method.
+
+Number of support samples. Table 5 lists the number of support samples for models of various sizes trained on the wikitext-2 dataset, showing consistency across scales.
+
+We further replicate most of results using 774M model with the wikitext-2 dataset. Following are some key observations.
+
+POS tags of support samples. Figure 11 shows that for 774M model,  $82\%$  of verbs are support samples while the proportion is only  $16\%$  for punc
+
+![](images/9cccc50c9e015a71a77987b00fc14c84930be3055b0c591464e6e1ec7b3d1a26.jpg)  
+Figure 11: Proportion of support samples in different POS tags of 774M model.
+
+tuations. This feature is consistent with the 117M model.
+
+Removing non-support samples. We conduct the experiment of removing non-support samples on the 774M model. As shown in Table 4, the conclusions(§4) still hold at this larger scale. For only training LM heads, we can remove non-support samples(around  $50\%$  of the samples) while keeping the same performances, with weighted sampling. Following the same reasoning(§4), we get for tuning heads, support samples are key, but some non-support samples are also necessary. For full model training, we can't simply remove non-support samples. Soft sampling outperforms random sampling when training LM heads, but the opposite is true for full model training, suggesting non-support sam
+
+<table><tr><td>Model Size</td><td>Eval Loss</td><td>#Support Samples(Prop.)</td></tr><tr><td>117M</td><td>5.05</td><td>1.29M (54%)</td></tr><tr><td>345M</td><td>5.01</td><td>1.28M (54%)</td></tr><tr><td>774M</td><td>5.02</td><td>1.13M (47%)</td></tr><tr><td>1.5B</td><td>5.13</td><td>1.26M (53%)</td></tr></table>
+
+Table 5: Evaluation loss and number of support samples for models of various sizes trained on the wikitext-2 dataset.
+
+plies are crucial for learning representations beyond the heads.

ACL/2025/On Support Samples of Next Word Prediction/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b3dcc1ff2168c4dc1a59df249fd97d5e4cc889e8a5fb092fadbb0b10e9154aad
+size 542447

ACL/2025/On Support Samples of Next Word Prediction/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1806267b64b554198bd55b915af21a3642bcd8607b0ba5ace0dd80a823bc28f0
+size 512816

ACL/2025/On Synthesizing Data for Context Attribution in Question Answering/0b5135dc-7731-47de-bc62-e3ae027200d3_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:96c7f99aa2dbdf206f69d44f5b0d4f9d3a80b3d185d84c727103e449bba08c21
+size 138812

ACL/2025/On Synthesizing Data for Context Attribution in Question Answering/0b5135dc-7731-47de-bc62-e3ae027200d3_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b9768de7f0db22d29b97e513aa36ed9a01e1287c4c987efd6f85113d15a05332
+size 177320

ACL/2025/On Synthesizing Data for Context Attribution in Question Answering/0b5135dc-7731-47de-bc62-e3ae027200d3_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:93eafed095433e8ad31fee660ca70b0f5b66472bd119224c05482ae55d6b2265
+size 834547

ACL/2025/On Synthesizing Data for Context Attribution in Question Answering/full.md

@@ -0,0 +1,560 @@
+# On Synthesizing Data for Context Attribution in Question Answering
+
+Gorjan Radevski $^{5,6*}$ , Kiril Gashteovski $^{1,8*}$ , Shahbaz Syed $^{1}$ , Christopher Malon $^{2}$ , Sebastien Nicolas $^{1}$ , Chia-Chien Hung $^{1}$ , Timo Sztyler $^{1}$ , Verena Heu $\beta$ er $^{1}$ , Wiem Ben Rim $^{4}$ , Masafumi Enomoto $^{3}$ , Kunihiro Takeoka $^{3}$ , Masafumi Oyamada $^{3}$ , Goran Glavaš $^{7}$ , Carolin Lawrence $^{1}$
+
+$^{1}$ NEC Laboratories Europe, Germany  $^{2}$ NEC Laboratories America, USA
+
+$^{3}$ NEC Corporation, Japan  $^{4}$ University College London, UK
+
+$^{5}$ KU Leuven, Belgium  $^{6}$ Epimind, Belgium
+
+<sup>7</sup>Center for Artificial Intelligence and Data Science, University of Würzburg, Germany
+
+$^{8}$ CAIR, Ss. Cyril and Methodius University of Skopje, North Macedonia
+
+# Abstract
+
+Question Answering (QA) accounts for a significant portion of LLM usage "in the wild". However, LLMs sometimes produce false or misleading responses, also known as hallucinations. Therefore, grounding the generated answers in contextually provided information—i.e., providing evidence for the generated text—is paramount for LLMs' trustworthiness. Providing this information is the task of context attribution. In this paper, we systematically study LLM-based approaches for this task, namely we investigate (i) zero-shot inference, (ii) LLM ensembling, and (iii) fine-tuning of small LMs on synthetic data generated by larger LLMs. Our key contribution is SYNQA: a novel generative strategy for synthesizing context attribution data. Given selected context sentences, an LLM generates QA pairs that are supported by these sentences. This leverages LLMs' natural strengths in text generation while ensuring clear attribution paths in the synthetic training data. We show that the attribution data synthesized via SYNQA is highly effective for fine-tuning small LMs for context attribution in different QA tasks and domains. Finally, with a user study, we validate the usefulness of small, efficient LMs (fine-tuned on synthetic data from SYNQA) in context attribution for QA.
+
+# 1 Introduction
+
+Large Language Models (LLMs) have become ubiquitous, with Question Answering (QA) as their most common use case (Trippas et al., 2024). However, LLMs have a tendency to hallucinate: they generate content that is factually incorrect w.r.t. a previously provided reference text. This poses the need for context attribution methods that create links between the answer and different relevant parts of the (potentially large) reference text; for an illustration of the task, see Figure 1.
+
+Michael Jordan is a former professional basketball player. He played 15 seasons in the National Basketball Association (NBA) between 1984 and 2003. He won 6 NBA championships with the Chicago Bulls.
+
+![](images/7997cadac65afea542df3990b3d82856a42b70ba00817aea13a14b8d140e2982.jpg)  
+Question Answering (given context)  
+Michael Jordan
+
+Question (human): Where is the team Michael Jordan, who is on the team?
+
+![](images/fae86e80b5b50510f71b119a479c0c022110605c3d03c4a204f2d2506ca1bfb4.jpg)  
+Attribution sentences
+
+The Chicago Bulls are an American professional basketball team based in Chicago, Illinois. The team was founded on January 16, 1966. The Bulls play their home games at the United Center.
+
+![](images/925e64536d527e2d54d982708cb7e07e22d63b158359bfd74e937f829df4994d.jpg)  
+Chicago Bulls
+
+Answer (LLM): In Chicago.
+
+(45) The Chicago Bulls are an American professional basketball club in the United States.
+
+![](images/a10c6553f97ca6c557edb6dfe5049e9c9d633ef59e9a46a8ae7517d926f80522.jpg)  
+Figure 1: Post-hoc context attribution: Given a question, an LLM-generated answer, and context (from human input or retrieval), the model identifies supporting sentences within the context. Our user study (§3.3.5) shows that presenting these supporting sentences helps users verify LLM answers more quickly and accurately.
+
+For these reasons, reliable and efficient context attribution is instrumental in manually verifying the factuality of LLM-generated content. By conducting a user study, Slobodkin et al. (2024) report two important findings in this respect: (1) attribution models reduce the human workload in a fact-checking task by as much as  $50\%$ ; and (2) sentence level granularity—i.e., grounding the answers in one or more relevant sentences in the reference context—is the most efficient granularity level for manual fact-checking of LLM-generated answers.
+
+Given the task importance, recent context attribution research spans text summarization (Krishna et al., 2023; Ernst et al., 2024), citation attribution (Gao et al., 2023b; Huang et al., 2024b), and question answering (Phukan et al., 2024; Cohen-Wang et al., 2024). However, the solutions rely on document- or paragraph-level evidence, which comes with the following limitations: (1) the user still has to read the (potentially long) document(s) to verify the generated text; and (2) the LLM needs to correctly generate the reference output alongside answering the question correctly. In contrast, the
+
+post-hoc attribution methods perform the attribution after the LLM generates the answer. These models, however, either attribute to coarse-grained units of text (Nakano et al., 2021; Menick et al., 2022; Buchmann et al., 2024), or provide fine-grained attributions but their inference is computationally expensive (Cohen-Wang et al., 2024), hindering their adoption in practice.
+
+In this paper, we explore how LLMs can generate synthetic data for attribution fine-tuning, enabling accurate, sentence-level, and real-time efficient models. For data generation, we compare two approaches: (1) in the fairly straightforward attribution synthesis (SYN-ATT), we start with question-answer pairs from a reference text and prompt the LLM to identify the supporting sentences; (2) in our novel question-answer synthesis (SYNQA), we use Wikipedia sentences and prompt the LLM to generate a question-answer pair that is fully supported by these sentences. Given the generated data, we fine-tune smaller, more efficient context attribution models and compare their performance.
+
+Through extensive evaluation encompassing six datasets and two real-world scenarios (attribution for single-turn questions: i.e., a single question and single answer, and for dialogue questions: i.e., as part of a conversation), we demonstrate that models trained on synthetic data generated by SYNQA: ① Outperform zero-shot LLMs that are orders of magnitude larger, while maintaining real-time inference capabilities ( $\S$ 3.3.1); ② Achieve competitive performance on in-domain tasks and superior generalization to out-of-domain datasets compared to models trained on gold data ( $\S$ 3.3.2); ③ Successfully handle dialogue-based attribution without requiring in-domain training data ( $\S$ 3.3.3); ④ Show consistent performance improvements as synthetic training data increases ( $\S$ 3.3.4); ⑤ Significantly improve users' speed and accuracy in verifying LLM-generated answers ( $\S$ 3.3.5).
+
+These findings suggest that SYNQA reduces dependence on large human-labeled datasets while improving context attribution robustness. Our user study further validates the practical utility of fine-tuned small models in real-world question-answering applications, demonstrating their effectiveness in diverse settings. Overall, these results highlight the viability of scalable, data-efficient context attribution techniques, paving the way for more interpretable and trustworthy AI systems.
+
+# 2 Synthesizing Attribution Data
+
+Context attribution identifies which parts of a reference text support a given question-answer pair (Rashkin et al., 2023). Formally, given a question  $q$ , its answer  $a$ , and a context text  $c$  consisting of sentences  $s_1, \ldots, s_n$ , the task is to identify the subset of sentences  $S \subseteq c$  that fully support the answer  $a$  to question  $q$ . To train efficient attribution models without requiring expensive human annotations, we explore synthetic data generation approaches using LLMs. We implement two methods for synthetic data generation. The baseline method (SYN-ATT) is discriminative: given existing question-answer pairs and their context, an LLM identifies supporting sentences (i.e., synthetic context attributions), which are then used to train a smaller attribution model using knowledge distillation. Our proposed method (SYNQA) takes a generative approach: given selected context sentences, an LLM generates question-answer pairs that are fully supported by these sentences. This approach better leverages LLMs' natural strengths in text generation while ensuring clear attribution paths in the synthetic training data.
+
+# 2.1 SYNQA: Generative Synthetic Data Generation Method
+
+SYNQA consists of three parts: context selection, QA generation, and distractors mining (for an illustration of the method, see Figure 2). In what follows, we describe each part in detail.
+
+Context Collection. We use Wikipedia as our data source, as each article consists of sentences about a coherent and connected topic, with two collection strategies. In the first, we select individual Wikipedia articles for dialogue-centric generation and use their sentences as context. In the second, for multi-hop reasoning, we identify sentences containing Wikipedia links and follow these links to create "hops" between articles, limiting to a maximum of two paths to maintain semantic coherence, while enabling more complex reasoning patterns. See Appendix B.1 for more details.
+
+Question-Answer Generation. Given the set of contexts, an LLM (Llama 70B in our implementation) can now generate question-answer pairs. For single articles, we prompt the model to generate a set of dialogue-centric question-answer pairs, where questions build upon the previous context. For linked articles, we prompt the model to generate questions that necessitate connecting infor
+
+![](images/3bbfc3d35b6a01dbadc37516b381bba7e20b70928af0ceb459f4ee8632b8ae5d.jpg)
+
+![](images/1c45c410f5540f1b5a93fcdeacd598d2d56ebf952b81ee0a4b22607f86a6a279.jpg)  
+Figure 2: Top: The SYN-ATT baseline method for synthetic attribution data generation. Given context and question-answer pairs, we prompt an LLM to identify supporting sentences, which are then used to train a smaller attribution model. However, this discriminative approach may yield noisy training data as LLMs are less suited for classification tasks (see §3.3.1). Bottom: The SYNQA data generation pipeline leverages LLMs' generative strengths through four steps: (1) collection of Wikipedia articles as source data; (2) extraction of context attributions by creating chains of sentences that form hops between articles; (3) generation of QA pairs by prompting an LLM with only these context attribution sentences; (4) compilation of the final training samples, each containing the generated QA pair, its context attributions, and the original articles enriched with related distractors.
+
+mation across the articles, encouraging multi-hop reasoning. Importantly, we provide the LLM with the complete multi-hop reasoning chain as ground truth attribution sentences and ask it to generate question-answer pairs that can only be answered using this evidence. This yields multi-hop samples requiring integration of information across documents. We provide more details in Appendix B.2, and the prompts we use for generating the synthetic data in Appendix D.
+
+Distractors Mining. To make the attribution task more realistic, we augment each sample with distractor articles. With E5 (Wang et al., 2022), we embed each Wikipedia article in our collection. For each article in the training sample, we randomly select up to three distractors with the highest semantic similarity to the source articles. These distractors share thematic elements with the source articles, but lack information to answer the questions. Since Wikipedia articles are unique within a single dump,
+
+the source article itself is never included among the distractors. The only scenario where partial information might exist would be if a question addresses a general fact that appears in multiple related articles—a rare occurrence given Wikipedia's article structure and our question generation process. See Appendix B.3 for more details.
+
+# 2.2 Advantages of SYNQA
+
+The SYNQA approach has three key advantages: (1) it leverages LLMs' strength in generation rather than classification; (2) creates diverse training samples requiring both dialogue understanding and multi-hop reasoning; and (3) ensures generated questions have clear attribution paths since they are derived from specific context sentences. By generating both entity-centric and dialogue-centric samples, SYNQA produces training data that reflects the variety of real-world QA scenarios, helping models develop robust attribution capabilities,
+
+which our experiments demonstrate to generalize across different contexts and domains.
+
+# 3 Experimental Study
+
+We conduct a comprehensive evaluation across multiple aspects: comparison with zero-shot LLMs, comparison with models trained on gold attribution data, and generalization to dialogue settings. With our experiments, we shed light on the performance and practical utility of our approach.
+
+# 3.1 Experimental Setting
+
+We evaluate model performance using precision (P), recall (R), and F1 score. For each sentence in the LLM's output, the context attribution models identify the set of context sentences that support that output sentence. Precision measures the proportion of predicted attributions that are correct, while recall measures the proportion of ground truth attributions that are successfully identified.
+
+For a fair and comprehensive evaluation, we train all models with a single pass over the training data unless stated otherwise, referring to this setup as 1P when needed. For a more controlled comparison, some experiments limit the number of training samples each model encounters. Since the synthetic dataset contains approximately 1.0M samples, we allow models to observe an equivalent number of samples from the gold training set, ensuring comparable exposure to models trained on data from SYNQA. We refer to this setting as 1M when necessary. For all models, we fine-tune only the LoRA parameters (alpha=64, rank=32) using a fixed learning rate of 1e-5 and a weight decay of 1e-3.
+
+In-domain datasets: We use SQuAD (Rajpurkar et al., 2016) and HotpotQA (Yang et al., 2018) as our primary in-domain benchmarks. SQuAD provides clear sentence-level evidence for answering questions, serving as a strong baseline for direct attribution. HotpotQA introduces multi-hop reasoning, requiring models to link information across multiple sentences (sometimes from different articles) to identify the correct evidence chain. Additionally, HotpotQA includes distractor documents—closely related yet incorrect sources—posing a more challenging but realistic setting for evaluating attribution performance.
+
+Out-of-domain datasets: To assess generalization beyond the training distribution, we evalu
+
+ate models on QuAC (Choi et al., 2018), CoQA (Reddy et al., 2018), OR-QuAC (Qu et al., 2020), and DoQA (Campos et al., 2020). These datasets present conversational QA scenarios that differ from SQuAD and HotpotQA. Specifically, QuAC and CoQA introduce multi-turn dialogue structures with coreferences, challenging models to track context across multiple turns. This conversational nature creates a methodological challenge: while these datasets are valuable for evaluating dialogue-based attribution, their reliance on conversation history makes direct comparison with models trained on single-turn QA datasets impossible.
+
+To enable comprehensive evaluation across dialogue QA and single-turn QA, we create two versions from Quac and CoQA: (i) a rephrased version using Llama 70B (Dubey et al., 2024) that converts questions into a standalone format for fair comparison with models trained on single-turn context attribution (simplified by “-ST”), and (ii) the original (unchanged) version for assessing dialogue-based attribution. See Appendix C for examples.
+
+DoQA extends this challenge further by incorporating domain-specific dialogues (cooking, travel, and movies), thus testing the models' adaptability to specialized contexts. OR-QuAC includes context-independent rewrites of the dialogue questions, such that they can be posed in isolation of prior context (i.e., single-turn QA). This enables us to test the models on their capabilities in both single-turn QA and dialogue QA settings.
+
+# 3.2 Methods
+
+We compare models trained with data from SYNQA against several baselines, including sentence encoder-based models, zero-shot instruction-tuned LLMs, and models trained on synthetic and gold context attribution data. Specifically, we experiment with the following methods:
+
+Sentence-Encoders: We embed each sentence in the context along with the question-answer pair, and select attribution sentences (from the context) based on the cosine similarity with a fixed threshold, tuned on a small validation set.
+
+Zero-shot LLMs: We evaluate various instruction-tuned LLMs in a zero-shot manner, as such models have been shown to perform well across a range of NLP tasks (Shu et al., 2023; Zhang et al., 2023). During inference, we provide an instruction template describing the task to the LLM (see Appendix E for details).
+
+Ensembles of LLMs: We aggregate the predictions of multiple LLMs through majority voting, selecting attribution sentences that receive consensus from at least  $50\%$  of the ensemble. In our experiments, we use Llama8B (Dubey et al., 2024), Mistral7B, and Mistral-Nemo12B (Jiang et al., 2023) as the ensemble constituents.
+
+Models trained on in-domain gold data: Finetuning on gold-labeled attribution data provides an upper bound on in-domain performance, helping us assess how well synthetic training data generalizes.
+
+SYN-ATT: SYN-ATT generates synthetic training data by prompting multiple LLMs to perform context attribution in a discriminative manner, aggregating their outputs via majority voting, and training a smaller model on the resulting dataset. To make it a stronger baseline against SYNQA, we give the training data of SQuAD and HotpotQA (the context, questions, and answers) to the LLMs and ask them to perform context attribution (note that we do not use the gold attribution). Finally, we train a model on the generated synthetic data—the context, questions, and answers from the training dataset, and the synthetic context attribution links.
+
+SYNQA: We train models using synthetic data generated by our proposed method (SYNQA). Importantly, we ensure that the models are not exposed to any part of the evaluation data.2
+
+# 3.3 Results and Discussion
+
+Evaluating our context attribution models requires a multifaceted approach, as performance is influenced by both the quality of training data and the model's ability to generalize beyond in-domain distributions. Therefore, we design our experiments to address five core questions: (i) How well do zero-shot LLMs perform on context attribution QA tasks ( $\S 3.3.1$ )? (ii) Can models trained on synthetic data generated by SYNQA exceed the performance of models trained on gold context attribution data ( $\S 3.3.2$ )? (iii) To what extent do models generalize to dialogue settings where in-domain training data is unavailable ( $\S 3.3.3$ )? (iv) How well do models scale in terms of synthetic data quantity generated by SYNQA ( $\S 3.3.4$ )? (v) How do improved context
+
+attributes impact the end users' speed and ability to verify questions answering outputs (§3.3.5)?
+
+# 3.3.1 Comparison to Zero-Shot Models
+
+In Table 1, we present the performance of zero-shot models and those trained without gold context attribution data. State-of-the-art sentence-encoder models (e.g., E5) perform relatively poorly, consistent with prior findings (Cohen-Wang et al., 2024). In contrast, LLMs exhibit strong performance, with improvements correlating with model size. Ensemble multiple zero-shot LLMs leverages complementary strengths and further enhances performance, but makes the attribution more expensive.
+
+Models trained using SYNQA data outperform all zero-shot baselines except on CoQA. This exception is particularly noteworthy given the nature of CoQA: while both QuAC(-ST) and CoQA(-ST) are conversational QA datasets (different from SQuAD and HotpotQA), CoQA is derived from diverse sources. On HotpotQA, we observe that Llama 70B exhibits high precision  $(87.6\%)$ , indicating that it rarely selects irrelevant sentences as supporting evidence. However, Llama 1B trained with SYNQA data achieves even higher precision  $(89.6\%)$ , suggesting that SYNQA further refines evidence selection. More importantly, SYNQATrained models exhibit higher recall compared to zero-shot LLMs, indicating they retrieve more complete supporting evidence for the QA pairs. This pattern demonstrates that SYNQA enables models to identify more comprehensive supporting evidence while maintaining strong precision.
+
+We also tested models trained with the discriminative method SYN-ATT. These models significantly outperform their non-fine-tuned same size counterparts. However, as postulated, our generative approach SYNQA outperforms SYN-ATT significantly in all cases as per F1 score. Additionally, SYNQA surpasses zero-shot LLMs that are orders of magnitude larger, showing that we can train a model that is both more accurate and efficient.
+
+# 3.3.2 Comparison to Models Trained on Gold Attribution Data
+
+In Table 2, we compare models trained on synthetic and gold in-domain context attribution datasets. As expected, fine-tuning on in-domain gold datasets (SQuAD and HotpotQA) yields highly specialized models that perform well on in-domain data.
+
+<table><tr><td rowspan="2">Model</td><td rowspan="2">Training data</td><td colspan="3">SQuAD</td><td colspan="3">HotpotQA</td><td colspan="3">Quac-ST</td><td colspan="3">CoQA-ST</td></tr><tr><td>P</td><td>R</td><td>F1</td><td>P</td><td>R</td><td>F1</td><td>P</td><td>R</td><td>F1</td><td>P</td><td>R</td><td>F1</td></tr><tr><td colspan="14">Baselines</td></tr><tr><td>Random</td><td>-</td><td>19.8</td><td>15.4</td><td>17.3</td><td>4.8</td><td>15.2</td><td>7.3</td><td>5.2</td><td>15.1</td><td>7.7</td><td>7.3</td><td>15.1</td><td>9.9</td></tr><tr><td>E5 | 561M</td><td>Zero-shot</td><td>38.1</td><td>76.5</td><td>50.9</td><td>12.4</td><td>41.4</td><td>19.1</td><td>65.0</td><td>73.8</td><td>69.1</td><td>61.1</td><td>15.2</td><td>24.4</td></tr><tr><td>HF-SmolLM2 | 365M</td><td>Zero-shot</td><td>28.1</td><td>46.4</td><td>35.0</td><td>5.1</td><td>7.3</td><td>6.0</td><td>10.6</td><td>22.6</td><td>14.4</td><td>10.6</td><td>21.5</td><td>14.2</td></tr><tr><td>Llama | 1B</td><td>Zero-shot</td><td>37.5</td><td>62.0</td><td>46.7</td><td>5.3</td><td>28.1</td><td>8.9</td><td>8.8</td><td>65.4</td><td>15.4</td><td>11.9</td><td>52.8</td><td>19.4</td></tr><tr><td>Mistral | 7B</td><td>Zero-shot</td><td>71.5</td><td>94.4</td><td>81.4</td><td>42.9</td><td>42.7</td><td>42.8</td><td>63.2</td><td>88.6</td><td>73.8</td><td>59.0</td><td>72.2</td><td>64.9</td></tr><tr><td>Llama | 8B</td><td>Zero-shot</td><td>71.9</td><td>96.9</td><td>82.6</td><td>49.2</td><td>52.9</td><td>51.0</td><td>64.1</td><td>92.1</td><td>75.6</td><td>55.7</td><td>76.4</td><td>64.4</td></tr><tr><td>Mistral-NeMo | 12B</td><td>Zero-shot</td><td>89.5</td><td>94.5</td><td>91.8</td><td>46.4</td><td>47.3</td><td>46.8</td><td>81.8</td><td>85.3</td><td>83.5</td><td>79.0</td><td>67.2</td><td>72.6</td></tr><tr><td>Ensemble | 27B</td><td>Zero-shot</td><td>83.1</td><td>96.3</td><td>89.2</td><td>48.1</td><td>59.6</td><td>53.2</td><td>74.8</td><td>90.3</td><td>81.8</td><td>69.5</td><td>73.6</td><td>71.5</td></tr><tr><td>Llama | 70B</td><td>Zero-shot</td><td>95.3</td><td>95.6</td><td>95.5</td><td>87.6</td><td>37.5</td><td>52.5</td><td>89.7</td><td>87.8</td><td>88.7</td><td>87.5</td><td>73.3</td><td>79.8</td></tr><tr><td colspan="14">Baselines</td></tr><tr><td>Llama | 1B</td><td>SYN-ATT (1P)</td><td>89.8</td><td>96.5</td><td>93.0</td><td>50.6</td><td>58.6</td><td>54.3</td><td>64.9</td><td>91.5</td><td>75.9</td><td>53.1</td><td>75.5</td><td>62.3</td></tr><tr><td>Llama | 1B</td><td>SYN-ATT (1M)</td><td>84.3</td><td>96.9</td><td>90.2</td><td>54.4</td><td>58.0</td><td>56.1</td><td>63.4</td><td>92.4</td><td>75.2</td><td>52.5</td><td>77.5</td><td>62.6</td></tr><tr><td colspan="14">Ours</td></tr><tr><td>Llama | 1B</td><td>SYNQA</td><td>96.0</td><td>96.2</td><td>96.1</td><td>89.6</td><td>69.4</td><td>78.2</td><td>93.3</td><td>89.1</td><td>91.1</td><td>82.3</td><td>68.5</td><td>74.8</td></tr></table>
+
+Table 1: Performance comparison between zero-shot models and models trained with synthetic data. While larger zero-shot language models perform well, our smaller SYNQA model achieves the highest F1 scores across all tasks. Bold indicates the best-performing method; underlined indicates our method when it ranks second.  $1\mathrm{P} =$  models trained with a single pass over training data;  $1\mathrm{M} =$  models trained with 1M samples (matched to SYNQA data size).  
+
+<table><tr><td rowspan="3">Model</td><td rowspan="3">Training data</td><td colspan="6">In-Domain</td><td colspan="6">Out-of-Domain</td></tr><tr><td colspan="3">SQuAD</td><td colspan="3">HotpotQA</td><td colspan="3">QuAC-ST</td><td colspan="3">CoQA-ST</td></tr><tr><td>P</td><td>R</td><td>F1</td><td>P</td><td>R</td><td>F1</td><td>P</td><td>R</td><td>F1</td><td>P</td><td>R</td><td>F1</td></tr><tr><td colspan="14">Baselines</td></tr><tr><td>Llama | 1B</td><td>Zero-shot</td><td>37.5</td><td>62.0</td><td>46.7</td><td>5.3</td><td>28.1</td><td>8.9</td><td>8.8</td><td>65.4</td><td>15.4</td><td>11.9</td><td>52.8</td><td>19.4</td></tr><tr><td>Llama | 1B</td><td>SQuAD (1P)</td><td>98.4</td><td>98.4</td><td>98.4</td><td>48.7</td><td>20.0</td><td>28.4</td><td>92.6</td><td>85.8</td><td>89.0</td><td>79.9</td><td>64.3</td><td>71.2</td></tr><tr><td>Llama | 1B</td><td>HotpotQA (1P)</td><td>41.3</td><td>87.3</td><td>56.0</td><td>87.5</td><td>79.9</td><td>83.5</td><td>45.2</td><td>89.9</td><td>60.1</td><td>41.0</td><td>70.9</td><td>52.0</td></tr><tr><td>Llama | 1B</td><td>SQuAD &amp; HotpotQA (1P)</td><td>98.3</td><td>98.3</td><td>98.3</td><td>89.7</td><td>78.9</td><td>84.0</td><td>90.4</td><td>90.0</td><td>90.2</td><td>83.1</td><td>68.0</td><td>74.8</td></tr><tr><td>Llama | 1B</td><td>SQuAD &amp; HotpotQA (1M)</td><td>98.3</td><td>98.4</td><td>98.3</td><td>87.0</td><td>85.2</td><td>86.1</td><td>84.0</td><td>89.2</td><td>86.6</td><td>79.2</td><td>66.4</td><td>72.2</td></tr><tr><td colspan="14">Ours</td></tr><tr><td>Llama | 1B</td><td>SYNQA</td><td>96.0</td><td>96.2</td><td>96.1</td><td>89.6</td><td>69.4</td><td>78.2</td><td>93.3</td><td>89.1</td><td>91.1</td><td>82.3</td><td>68.5</td><td>74.8</td></tr><tr><td>Llama | 1B</td><td>SYNQA &amp; SQuAD &amp; HotpotQA</td><td>98.2</td><td>98.3</td><td>98.2</td><td>89.3</td><td>82.4</td><td>85.8</td><td>94.5</td><td>92.7</td><td>93.6</td><td>85.5</td><td>71.0</td><td>77.6</td></tr></table>
+
+Table 2: Performance comparison of models fine-tuned on synthetic versus gold in-domain data. Our SYNQA approach demonstrates superior generalization while maintaining competitive performance on in-domain tasks. Bold indicates the best-performing method; underlined indicates our method when it ranks second.  $1\mathrm{P} =$  models trained with a single pass over training data;  $1\mathrm{M} =$  models trained with 1M samples (matched to SYNQA data size).
+
+However, models trained on data obtained by SYNQA exhibit competitive performance on in-domain tasks and consistently surpass in-domain-trained models on out-of-domain datasets. Finetuning models on in-domain data (SQuAD and HotpotQA, in addition to SYNQA) further improves recall, particularly on HotpotQA. Specifically, recall improves from  $69.4\%$  to  $82.4\%$  when adding in-domain data to models trained with SYNQA. We observe a similar trend for QuAC-ST and CoQA-ST: SYNQA-trained models already exhibit high precision, but recall is further enhanced when adding in-domain data. This suggests that while SYNQA effectively teaches the model to precisely attribute evidence, fine-tuning with domain
+
+specific data allows it to capture a broader set of relevant evidence. Such out-of-domain generalization is crucial for practical deployments, where models must handle diverse, unseen contexts that often differ substantially from their training data.
+
+# 3.3.3 Comparison to Zero-Shot and Fine-Tuned Models in Dialogue Settings
+
+We evaluate dialogue context attribution, for which we do not use any gold in-domain training data (Table 3).4 Here, models must handle follow-up questions that rely on previous turns, often involving coreferences and other dialogue-specific
+
+<table><tr><td rowspan="3">Model</td><td rowspan="3">Training data</td><td colspan="12">Out-of-Domain</td></tr><tr><td colspan="3">QuAC</td><td colspan="3">CoQA</td><td colspan="3">OR-QuAC</td><td colspan="3">DoQA</td></tr><tr><td>P</td><td>R</td><td>F1</td><td>P</td><td>R</td><td>F1</td><td>P</td><td>R</td><td>F1</td><td>P</td><td>R</td><td>F1</td></tr><tr><td colspan="14">Baselines</td></tr><tr><td>Llama | 1B</td><td>Zero-shot</td><td>30.8</td><td>45.5</td><td>36.8</td><td>39.4</td><td>37.9</td><td>38.6</td><td>33.0</td><td>46.6</td><td>38.6</td><td>12.2</td><td>22.6</td><td>15.9</td></tr><tr><td>Mistral | 7B</td><td>Zero-shot</td><td>76.6</td><td>81.8</td><td>79.1</td><td>67.6</td><td>61.3</td><td>64.3</td><td>82.5</td><td>85.1</td><td>83.8</td><td>74.9</td><td>77.9</td><td>76.4</td></tr><tr><td>Llama | 8B</td><td>Zero-shot</td><td>84.7</td><td>88.8</td><td>86.7</td><td>79.3</td><td>72.0</td><td>75.5</td><td>88.0</td><td>91.3</td><td>89.6</td><td>77.9</td><td>91.4</td><td>84.1</td></tr><tr><td>Mistral-NeMo | 12B</td><td>Zero-shot</td><td>85.7</td><td>85.4</td><td>85.5</td><td>81.9</td><td>68.4</td><td>74.5</td><td>88.9</td><td>88.8</td><td>88.8</td><td>86.0</td><td>84.2</td><td>85.1</td></tr><tr><td>Llama | 70B</td><td>Zero-shot</td><td>88.5</td><td>87.7</td><td>88.1</td><td>88.3</td><td>74.9</td><td>81.1</td><td>81.7</td><td>86.3</td><td>83.9</td><td>85.2</td><td>82.0</td><td>83.5</td></tr><tr><td colspan="14">Baselines</td></tr><tr><td>Llama | 1B</td><td>SQuAD &amp; HotpotQA (1P)</td><td>71.3</td><td>66.8</td><td>69.0</td><td>79.0</td><td>64.2</td><td>70.8</td><td>61.6</td><td>57.5</td><td>59.5</td><td>67.4</td><td>57.8</td><td>62.2</td></tr><tr><td>Llama | 1B</td><td>SQuAD &amp; HotpotQA (1M)</td><td>52.6</td><td>49.3</td><td>50.9</td><td>61.2</td><td>50.2</td><td>55.2</td><td>48.5</td><td>44.6</td><td>46.5</td><td>53.2</td><td>49.1</td><td>51.1</td></tr><tr><td colspan="14">Ours</td></tr><tr><td>Llama | 1B</td><td>SYNQA</td><td>91.3</td><td>91.4</td><td>91.3</td><td>81.7</td><td>71.4</td><td>76.2</td><td>92.6</td><td>95.3</td><td>94.0</td><td>86.3</td><td>94.5</td><td>90.2</td></tr><tr><td>Llama | 1B</td><td>SYNQA &amp; SQuAD &amp; HotpotQA</td><td>91.1</td><td>92.2</td><td>91.7</td><td>82.3</td><td>73.2</td><td>77.5</td><td>90.3</td><td>96.4</td><td>93.2</td><td>85.1</td><td>96.0</td><td>90.2</td></tr></table>
+
+Table 3: Context attribution on QuAC, CoQA, OR-Quac, and DoQA (dialogue data); all datasets are out-of-domain (i.e., we do not use the training sets). Our SYNQA models outperform fine-tuned and larger zero-shot models, despite their size advantage. Bold denotes best method, underline our method when second best. 1P: models trained with a single pass over the training data. 1M: models trained with 1M samples (matched to SYNQA data size).
+
+complexities. Since all dialogue datasets can be considered out-of-domain compared to SQuAD and HotpotQA (dialogue vs. single-turn; different sources—e.g., DoQA is derived from StackExchange while SQuAD and HotpotQA are derived from Wikipedia), the patterns we observe differ in some cases from the single-turn results. As expected, zero-shot LLMs exhibit a strong size-performance correlation, with larger models consistently outperforming smaller ones—even those fine-tuned on single-turn question-answer attribution (trained on gold SQuAD and HotpotQA). However, fine-tuning smaller models with SYNQA data leads to superior performance, surpassing both their fine-tuned counterparts and much larger zero-shot LMs. Unlike Table 2, training on SQuAD and HotpotQA does not lead to consistent improvement in dialogue settings. Since these datasets consist of single-turn QA, they are less beneficial for improving performance on multi-turn QA data. As with the single-turn results, zero-shot LLMs already exhibit high precision, but models using SYNQA data either match or surpass them while improving recall, further suggesting that SYNQA enhances the model's ability to extract a more complete set of supporting evidence, even in dialogue settings.
+
+# 3.3.4 Scaling Trends and Generalization Performance
+
+Fig. 3a shows F1 scores averaged across datasets, with model size on the x-axis and performance on the y-axis. Models trained on SYNQA-generated data significantly outperform their baseline zero-shot counterparts, while also achieving superior
+
+performance compared to zero-shot LLMs that are orders of magnitude larger. This shows our method is highly data-efficient, enabling small models to close the gap with much larger counterparts.
+
+In Figure 3b, we analyze model performance as the quantity of synthetic training data increases, reporting F1 scores separately for in-domain and out-of-domain datasets. As we scale data quantity, performance improves consistently across datasets for single-turn context attribution. This trend highlights the scalability of our approach, indicating that further gains can be achieved by increasing the quantity of the synthetic data.
+
+# 3.3.5 User Study: SYNQA increases efficiency and accuracy assessment
+
+We conducted a user study to evaluate the efficiency and accuracy of verifying the correctness of LLM-generated answers using context attribution. Our hypothesis is that higher-quality context attributions, visualized to guide users, facilitate faster and more accurate verification of LLM outputs. Specifically, in each trial, we presented users with a question, a generated answer, and relevant context, along with attributions visualized as highlights. Their task was to leverage these attributions to judge if the answer was correct w.r.t. a provided context. See Figure 5 in Appendix F.
+
+The study compares three scenarios: (i) No Alignment: a baseline condition without context attributions, requiring users to manually read and verify the answer against the entire context; (ii) Llama 1B (Zero-shot): context attributions
+
+![](images/7f39932b653e0c7a18e03e13ad91ba64955a4ecbe7255d9a470ca65ee4f9c014.jpg)  
+(a) Model performance vs. size.
+
+![](images/3627c4b2184a18fa8c93b050af0f10ca8814b82c9916946502a0f4312d8318b4.jpg)  
+(b) F1 score vs. training data size.
+
+generated by the Llama 1B model were visualized; (iii) SYNQA: context attributions generated by our approach were visualized.
+
+We employed a within-subjects experimental design for our human evaluation (with 12 participants), ensuring that the same participants evaluate all the aforementioned alignment scenarios, thus requiring fewer participants for reliable results (Greenwald, 1976). However, this can be susceptible to learning effects where participants perform better in later scenarios, because they learned the task from previous examples. To mitigate this, we counterbalanced the scenario order using a Latin Square design (Belz and Kow, 2010; Bradley, 1958), where each alignment scenario appears in each position an equal number of times across all participants. Finally, we randomized the example order within each scenario per participant. For each example, we measured: verification time (seconds from display to judgment submission) and verification accuracy (binary correct/incorrect judgment). Results. We observed a clear trend in verification performance across the different attribution settings, with SYNQA demonstrating superior effectiveness (Fig. 5). SYNQA has the lowest average verification time per example (148.6 seconds), significantly faster than No Alignment (171.8 seconds) and attributions from Llama 1B (163.4 seconds). Concurrently, in terms of verification accuracy, SYNQA achieved the highest average accuracy  $(86.4\%)$ . While No Alignment  $(84.1\%)$  and Llama 1B  $(77.3\%)$  also yielded reasonable ac
+
+![](images/5145d837d9f13f71efa9f71e636535df7d68d0ab3c58efe5b2fda366c68de674.jpg)  
+Figure 3: Comparison of model performance and scalability. (a) Larger zero-shot models achieve good F1 scores, but our method SYNQA (based on Llama 1B) outperforms them while being orders of magnitude smaller. (b) Performance improves consistently with more SYNQA training data, highlighting its scalability.  
+Figure 4: Relationship between Evaluation Time (seconds) and Accuracy (\%) for three answer verification settings: Llama 1B (Zero-shot), No Alignment and SYNQA. SYNQA demonstrates the lowest evaluation time and highest accuracy, indicating its superior performance in facilitating efficient and accurate answer verification.
+
+curacy, attributions from SYNQA are clearly of higher quality helping users be more accurate.
+
+# 4 Related Work
+
+Work on context attribution for QA can be split into two categories: (1) in-line citation generation: LLMs are instructed to generate citations along with the generated answer; (2) post-hoc context attribution: perform the attribution after the LLM generates the answer. In the following, we outline these works and their differences from our work. For a more comprehensive discussion on related work, see Appendix A.
+
+# 4.1 In-line Citation Generation
+
+In this setup, researchers use LLMs to produce inline citations along with the generated text (Bohnet et al., 2022; Gao et al., 2023b; Huang et al., 2024b). This typically works on paragraph or document level. One line of work focuses on fine-tuning methods for tackling the problem (Gao et al., 2023b; Schimanski et al., 2024; Berchansky et al., 2024; Patel et al., 2024), while another line of work proposes synthetic data generation methods for fine-tuning such models (Huang et al., 2024a,b). Slobodkin et al. (2024) propose a fine-grained task, where the attributions are on sentence level, because such granularity is more useful to human end users. Since generating such in-line citations can result in producing completely made-up citations, Yue et al. (2023) propose a task that checks whether in-line generated citations from LLMs are attributable or not. Unlike such approaches, we focus on post-hoc context attributions, because this directly predicts a link to a factual source, and therefore avoiding the risk of making up the source.
+
+# 4.2 Post-hoc Context Attribution
+
+In post-hoc context attribution, the aim is to determine which parts of the context are attributable to an already answered question (Yang et al., 2018). There has been a significant amount of work on training models for the context attribution problem on sentence-level for multi-hop QA (Zhang et al., 2024; Ho et al., 2023; Yin et al., 2023; Fu et al., 2021; Tu et al., 2020; Fang et al., 2020). However, they do not investigate this problem in the context of LLMs. Moreover, the methods are constrained only to multi-hop QA, and are not tested on broader QA context, such as dialogue QA. In our work, we propose methods that use LLMs as data generators. This allows us to better generalize and cover multiple QA settings simultaneously, therefore better matching real-world needs.
+
+Another line of work focuses on coarse-level granularity and provides attributions either on paragraph level (Rashkin et al., 2023; Menick et al., 2022) or document level (Nakano et al., 2021; Gao et al., 2023a; Buchmann et al., 2024). However, in a user study Slobodkin et al. (2024) observe that such granularity level is not optimal for humans when manually fact-checking LLM-generated content. Their experiments suggest that sentence-level granularity is ideal for humans. This is why we adopt sentence-level granularity in our
+
+work, despite this being a harder task. On the other hand, there has been work that focuses on the other extreme: assigning context attributions on sub-sentence level (Cohen-Wang et al., 2024; Phukan et al., 2024; Ramu et al., 2024). While these approaches provide more granular attributions, they are computationally more expensive, which hinders their practical usability. Our work ensures that models can be run in real-time to make them practical for end users.
+
+# 5 Conclusion
+
+We investigated the task of context attribution in QA. We focused on approaches that enhance attribution performance without relying on prohibitive human annotations. Our proposed data synthesis strategy, SYNQA, enables the generation of high-quality synthetic attribution data, leading to substantial improvements in fine-tuned small models.
+
+Through extensive experiments on six datasets across single turn QA and dialogue QA attribution, we demonstrated that small models fine-tuned with SYNQA data (i) significantly outperform models trained on alternative synthetic attributions, (ii) exceed the performance of zero-shot LLMs that are orders of magnitude larger, and (iii) generalize better to out-of-domain distributions compared to models trained on gold in-domain data. These findings suggest that SYNQA reduces reliance on large-scale human-labeled datasets, while improving attribution robustness across diverse scenarios.
+
+Finally, our user study validates the practical utility of fine-tuned small models in real-world question-answering applications. These results highlight the viability of scalable, data-efficient context attribution techniques, paving the way for more interpretable and trustworthy AI systems.
+
+# Limitations
+
+While our work demonstrates the effectiveness of SYNQA for context attribution in question answering, we leave some important directions for future research. First, all models we train operate exclusively at the sentence level. Even though Slobodkin et al. (2024), through a user study, found that sentence-level granularity of context attribution QA is probably the best-suited granularity for manual verification of LLM output, it might not always be optimal granularity for attribution in other tasks. Namely, some context elements might be better captured at different levels: e.g., from individual
+
+phrases to multi-sentence passages—depending on the semantic structure of the text.
+
+Second, while we evaluated our approach on OR-QuAC, we have not fully explored context attribution in retrieval-augmented generation (RAG) settings with dialogue. This presents unique challenges since models must simultaneously handle dialogue-style questions and continuously evolving context. Future work should examine how attribution models can adapt when relevant context changes throughout a conversation.
+
+Third, we focused primarily on question answering, however, context attribution is valuable for many other natural language processing tasks, e.g., in text summarization, attributing summary sentences to source document segments could enhance transparency and fact-checking capabilities. Future research should examine how SYNQA's synthetic data generation approach can be adapted for different tasks, potentially revealing task-specific challenges and nuances.
+
+Fourth, our user study (§3.3.5), while providing valuable initial insights into the effectiveness of context attribution to help users verify the LLM model outputs in QA settings, was conducted with a limited sample of 12 participants. A larger-scale study with more participants would strengthen the statistical validity of our findings and potentially reveal more nuanced patterns. Future work should extend this evaluation to a more diverse and larger participant pool, ideally, including users with varying levels of domain expertise and familiarity with language model outputs.
+
+# Acknowledgements
+
+We thank Andreas Ripke for his support with hosting the LLMs used in our experiments. We also thank Dina Trajkovska for turning our messy ideas into beautiful figures that actually make sense.
+
+# References
+
+R. Anantha, Svitlana Vakulenko, Zhucheng Tu, S. Long-pre, Stephen G. Pulman, and Srinivas Chappidi. 2020. Open-domain question answering goes conversational via question rewriting. In North American Chapter of the Association for Computational Linguistics.  
+Anja Belz and Eric Kow. 2010. Comparing rating scales and preference judgements in language evaluation. In Proceedings of the 6th International Natural Language Generation Conference. Association for Computational Linguistics.  
+Moshe Berchansky, Daniel Fleischer, Moshe Wasserblat, and Peter Izsak. 2024. CoTAR: Chain-of-Thought Attribution Reasoning with Multi-level Granularity. In Conference on Empirical Methods in Natural Language Processing (EMNLP), page 236-246.  
+Bernd Bohnet, Vinh Q. Tran, Pat Verga, Roee Aharoni, Daniel Andor, Livio Baldini Soares, Massimiliano Ciaramita, Jacob Eisenstein, Kuzman Ganchev, Jonathan Herzig, Kai Hui, Tom Kwiatkowski, Ji Ma, Jianmo Ni, Tal Schuster, William Cohen W., Michael Collins, Dipanjan Das, Donald Metzler, Slav Petrov, and Kellie Webster. 2022. Attributed Question Answering: Evaluation and Modeling for Attributed Large Language Models. arXiv preprint arXiv:2212.08037.  
+Claudio Delli Bovi, Luca Telesca, and Roberto Navigli. 2015. Large-scale information extraction from textual definitions through deep syntactic and semantic analysis. Transactions of the Association for Computational Linguistics, 3:529-543.  
+James V Bradley. 1958. Complete Counterbalancing of Immediate Sequential Effects in a Latin Square Design. Journal of the American Statistical Association, 53(282):525-528.  
+Samuel Broscheit, Kiril Gashteovski, Yanjie Wang, and Rainer Gemulla. 2020. Can We Predict New Facts with Open Knowledge Graph Embeddings? A Benchmark for Open Link Prediction. In Association for Computational Linguistics (ACL), pages 2296-2308.  
+Jan Buchmann, Xiao Liu, and Iryna Gurevych. 2024. Attribute or Abstain: Large Language Models as Long Document Assistants. In Conference on Empirical Methods in Natural Language Processing (EMNLP), page 8113-8140.  
+Jon Ander Campos, Arantxa Otegi, Aitor Soroa, Jan Deriu, Mark Cieliebak, and Eneko Agirre. 2020. DoQA-Accessing Domain-Specific FAQs via Conversational QA. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7302-7314, Online. Association for Computational Linguistics.  
+Eunsol Choi, He He, Mohit Iyyer, Mark Yatskar, Wen tau Yih, Yejin Choi, Percy Liang, and Luke Zettle-moyer. 2018. QuAC: Question Answering in Context.
+
+In Conference on Empirical Methods in Natural Language Processing (EMNLP).  
+Benjamin Cohen-Wang, Harshay Shah, Kristian Georgiev, and Aleksander Mądry. 2024. ContextCite: Attributing Model Generation to Context. ArXiv, abs/2409.00729.  
+Preetam Prabhu Srikar Dammu, Himanshu Naidu, Mouly Dewan, YoungMin Kim, Tanya Roosta, Aman Chadha, and Chirag Shah. 2024. ClaimVer: Explainable Claim-Level Verification and Evidence Attribution of Text Through Knowledge Graphs. ArXiv, abs/2403.09724.  
+Anirban Dasgupta, Ravi Kumar, and Tamás Sarlós. 2011. Fast Locality-Sensitive Hashing. In Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD), pages 1073-1081.  
+Pradeep Dasigi, Kyle Lo, Iz Beltagy, Arman Cohan, Noah A. Smith, and Matt Gardner. 2021. A Dataset of Information-Seeking Questions and Answers Anchored in Research Papers. In North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT), pages 4599-4610.  
+Emily Dinan, Stephen Roller, Kurt Shuster, Angela Fan, Michael Auli, and Jason Weston. 2019. Wizard of Wikipedia: Knowledge-Powered Conversational agents. In Conference on Learning Representations (ICLR).  
+Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. 2024. The Llama 3 Herd of Models. arXiv preprint arXiv:2407.21783.  
+David Dukić, Kiril Gashteovski, Goran Glavaš, and Jan Snajder. 2024. Leveraging open information extraction for more robust domain transfer of event trigger detection. In Findings of the Association for Computational Linguistics: EACL 2024, pages 1197-1213.  
+Ori Ernst, Avi Caciularu, Ori Shapira, Ramakanth Pasunuru, Mohit Bansal, Jacob Goldberger, and Ido Dagan. 2022. Proposition-level Clustering for Multidocument Summarization. In Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT), pages 1765-1779.  
+Ori Ernst, Ori Shapira, Ramakanth Pasunuru, Michael Lepioshkin, Jacob Goldberger, Mohit Bansal, and Ido Dagan. 2021. Summary-source Proposition-level Alignment: Task, Datasets and Supervised Baseline. In Conference on Computational Natural Language Learning (CoNLL), pages 310-322.  
+Ori Ernst, Ori Shapira, Aviv Slobodkin, Sharon Adar, Mohit Bansal, Jacob Goldberger, Ran Levy, and Ido Dagan. 2024. The Power of Summary-Source Alignments. In Findings of the Association for Computational Linguistics (ACL Findings), pages 6527-6548.
+
+Yuwei Fang, Siqi Sun, Zhe Gan, Rohit Pillai, Shuo-hang Wang, and Jingjing Liu. 2020. Hierarchical Graph Network for Multi-hop Question Answering. In Empirical Methods in Natural Language Processing (EMNLP), pages 8823-8838.  
+Ruiliu Fu, Han Wang, Xuejun Zhang, Jun Zhou, and Yonghong Yan. 2021. Decomposing Complex Questions Makes Multi-Hop QA Easier and more Interpretable. In Conference on Empirical Methods in Natural Language Processing (EMNLP Findings), pages 169-180.  
+Luyu Gao, Zhuyun Dai, Panupong Pasupat, Anthony Chen, Arun Tejasvi Chaganty, Yicheng Fan, Vincent Zhao, N. Lao, Hongrae Lee, Da-Cheng Juan, and Kelvin Guu. 2023a. RARR: Researching and Revising What Language Models Say, Using Language Models. In Annual Meeting of the Association for Computational Linguistics (ACL), pages 16477-16508.  
+Tianyu Gao, Howard Yen, Jiatong Yu, and Danqi Chen. 2023b. Enabling Large Language Models to Generate Text with Citations. In Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6465-6488.  
+Kiril Gashteovski, Rainer Gemulla, Bhushan Kotnis, Sven Hertling, and Christian Meilicke. 2020. On Aligning OpenIE Extractions with Knowledge Bases: A Case Study. In Proceedings of the Workshop on Evaluation and Comparison of NLP Systems, Eval4NLP@ACL, pages 143-154.  
+Kiril Gashteovski, Sebastian Wanner, Sven Hertling, Samuel Broscheit, and Rainer Gemulla. 2019. OPIEC: An Open Information Extraction Corpus. In Conference on Automated Knowledge Base Construction (AKBC).  
+Anthony G Greenwald. 1976. Within-Subjects Designs: To Use or not to Use? Psychological Bulletin, 83(2):314.  
+Xanh Ho, Anh-Khoa Duong Nguyen, Saku Sugawara, and Akiko Aizawa. 2023. Analyzing the Effectiveness of the Underlying Reasoning Tasks in Multi-hop Question Answering. In Conference of the European Chapter of the Association for Computational Linguistics (EACL Findings), pages 1133-1150.  
+Lei Huang, Xiaocheng Feng, Weitao Ma, Yuxuan Gu, Weihong Zhong, Xiachong Feng, Weijiang Yu, Weihua Peng, Duyu Tang, Dandan Tu, and Bing Qin. 2024a. Learning Fine-Grained Grounded Citations for Attributed Large Language Models. In Findings of the Association for Computational Linguistics (ACL Findings), pages 14095-14113.  
+Lei Huang, Xiaocheng Feng, Weitao Ma, Liang Zhao, Yuchun Fan, Weihong Zhong, Dongliang Xu, Qing Yang, Hongtao Liu, and Bing Qin. 2024b. Advancing Large Language Model Attribution through Self-Improving. In Conference on Empirical Methods in Natural Language Processing (EMNLP).
+
+Filip Ilievski, Barbara Hammer, Frank van Harmelen, Benjamin Paassen, Sascha Saralajew, Ute Schmid, Michael Biehl, Marianna Bolognesi, Xin Luna Dong, Kiril Gashteovski, Giuseppe Marra Pascal Hitzler, Pasquale Minervini, Martin Mundt, Axel-Cyrille Ngonga Ngomo, Alessandro Oltramari, Gabriella Pasi, Zeynep G. Saribatur, Luciano Serafini, John Shawe-Taylor, Vered Shwartz, Gabriella Skitalinskaya, Clemens Stachl, Gido M. van de Ven, and Thomas Villmann. 2024. Aligning Generalisation Between Humans and Machines. arXiv preprint arXiv:2411.15626.  
+Albert Qiaochu Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de Las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, L'elio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, and William El Sayed. 2023. Mistral 7B. ArXiv, abs/2310.06825.  
+Bhushan Kotnis, Kiril Gashteovski, Julia Gastinger, Giuseppe Serra, Francesco Alesiani, Timo Sztyler, Ammar Shaker, Na Gong, Carolin Lawrence, and Zhao Xu. 2022a. Human-Centric Research for NLP: Towards a Definition and Guiding Questions. arXiv preprint arXiv:2207.04447.  
+Bhushan Kotnis, Kiril Gashteovski, and Carolin Lawrence. 2023. Open Information Extraction from Low Resource Languages. US Patent 11,741,318.  
+Bhushan Kotnis, Kiril Gashteovski, Daniel Rubio, Ammar Shaker, Vanessa Rodriguez-Tembras, Makoto Takamoto, Mathias Niepert, and Carolin Lawrence. 2022b. MILIE: Modular & Iterative Multilingual Open Information Extraction. In Association for Computational Linguistics (ACL), pages 6939-6950.  
+Kalpesh Krishna, Erin Bransom, Bailey Kuehl, Mohit Iyyer, Pradeep Dasigi, Arman Cohan, and Kyle Lo. 2023. LongEval: Guidelines for Human Evaluation of Faithfulness in Long-form Summarization. In Conference of the European Chapter of the Association for Computational Linguistics (EACL), pages 1642-1661.  
+Dongfang Li, Zetian Sun, Xinshuo Hu, Zhenyu Liu, Ziyang Chen, Baotian Hu, Aiguo Wu, and Min Zhang. 2023. A Survey of Large Language Models Attribution. ArXiv, abs/2311.03731.  
+Yifei Li, Xiang Yue, Zeyi Liao, and Huan Sun. 2024. AttributionBench: How Hard is Automatic Attribution Evaluation? In Annual Meeting of the Association for Computational Linguistics (ACL), pages 14919-14935.  
+Himanshu Maheshwari, Sambaran Bandyopadhyay, Aparna Garimella, and Anandhavelu Natarajan. 2024. Presentations are not always Linear! GNN Meets LLM for Document-to-Presentation Transformation with Attribution. In Conference on Empirical Methods in Natural Language Processing (EMNLP, Findings), pages 15948-15962.
+
+Jacob Menick, Maja Trebacz, Vladimir Mikulik, John Aslanides, Francis Song, Martin Chadwick, Mia Glaese, Susannah Young, Lucy Campbell-Gillingham, Geoffrey Irving, and Nat McAleese. 2022. Teaching Language Models to Support Answers with Verified Quotes. ArXiv, abs/2203.11147.  
+Reiichiro Nakano, Jacob Hilton, Suchir Balaji, Jeff Wu, Ouyang Long, Christina Kim, Christopher Hesse, Shantanu Jain, Vineet Kosaraju, William Saunders, Xu Jiang, Karl Cobbe, Tyna Eloundou, Gretchen Krueger, Kevin Button, Matthew Knight, Benjamin Chess, and John Schulman. 2021. WebGPT: Browser-assisted Question-answering with Human Feedback. ArXiv, abs/2112.09332.  
+Ramesh Nallapati, Bowen Zhou, Cícero Nogueira dos Santos, Caglar Gúlçehre, and Bing Xiang. 2016. Abstractive Text Summarization using Sequence-to-sequence RNNs and Beyond. In Conference on Computational Natural Language Learning (CoNLL), pages 280-290.  
+Federico Nanni, Jingyi Zhang, Ferdinand Betz, and Kiril Gashteovski. 2019. EAL: A Toolkit and Dataset for Entity-Aspect Linking. In 2019 ACM/IEEE Joint Conference on Digital Libraries (JCDL), pages 430-431. IEEE.  
+Jianmo Ni, Chen Qu, Jing Lu, Zhuyun Dai, Gustavo Hernández Abrego, Ji Ma, Vincent Y Zhao, Yi Luan, Keith B Hall, Ming-Wei Chang, et al. 2022. Large Dual Encoders Are Generalizable Retrievers. In Empirical Methods in Natural Language Processing (EMNLP), pages 9844–9855.  
+Ankur Parikh, Xuezhi Wang, Sebastian Gehrmann, Manaal Faruqui, Bhuwan Dhingra, Diyi Yang, and Dipanjan Das. 2020. ToTTo: A controlled table-to-text generation dataset. In Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1173-1186.  
+Nilay Patel, Shivashankar Subramanian, Siddhant Garg, Pratyay Banerjee, and Amita Misra. 2024. Towards improved Multi-source Attribution for Long-form Answer Generation. In North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT), pages 3906-3919.  
+Anirudh Phukan, Shwetha Somasundaram, Apoorv Saxena, Koustava Goswami, and Balaji Vasan Srinivasan. 2024. Peering into the Mind of Language Models: An Approach for Attribution in Contextual Question Answering. In Annual Meeting of the Association for Computational Linguistics (ACL), pages 11481-11495.  
+Jirui Qi, Gabriele Sarti, Raquel Fern'andez, and Arianna Bisazza. 2024. Model Internals-based Answer Attribution for Trustworthy Retrieval-Augmented Generation. In Conference on Empirical Methods in Natural Language Processing (EMNLP), page 6037-6053.
+
+Chen Qu, Liu Yang, Cen Chen, Minghui Qiu, W Bruce Croft, and Mohit Iyyer. 2020. Open-Retrieval Conversational Question Answering. In Proceedings of the International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR), pages 539-548.  
+Gorjan Radevski, Kiril Gashteovski, Chia-Chien Hung, Carolin Lawrence, and Goran Glavaš. 2023. Linking Surface Facts to Large-Scale Knowledge Graphs. In Empirical Methods in Natural Language Processing (EMNLP), pages 7189-7207.  
+Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. SQuAD: 100,000+ Questions for Machine Comprehension of Text. In Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2383-2392.  
+Pritika Ramu, Koustava Goswami, Apoorv Saxena, and Balaji Vasan Srinivavsan. 2024. Enhancing Post-Hoc Attributions in Long Document Comprehension via Coarse Grained Answer Decomposition. In Empirical Methods in Natural Language Processing (EMNLP), pages 17790-17806.  
+Hannah Rashkin, Vitaly Nikolaev, Matthew Lamm, Lora Aroyo, Michael Collins, Dipanjan Das, Slav Petrov, Gaurav Singh Tomar, Iulia Turc, and David Reitter. 2023. Measuring Attribution in Natural Language Generation Models. Computational Linguistics, 49(4):777-840.  
+Siva Reddy, Danqi Chen, and Christopher D. Manning. 2018. CoQA: A Conversational Question Answering Challenge. Transactions of the Association for Computational Linguistics (TACL), 7:249-266.  
+Wiem Ben Rim, Ammar Shaker, Zhao Xu, Kiril Gashteovski, Bhushan Kotnis, Carolin Lawrence, Jürgen Quittek, and Sascha Saralajew. 2024. A Human-Centric Assessment of the Usefulness of Attribution Methods in Computer Vision. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases (ECML/PKDD), pages 20-37. Springer.  
+Tobias Schimanski, Jingwei Ni, Mathias Kraus, Elliott Ash, and Markus Leippold. 2024. Towards Faithful and Robust LLM Specialists for Evidence-Based Question-Answering. In Annual Meeting of the Association for Computational Linguistics (ACL), pages 1913-1931.  
+Manli Shu, Jiongxiao Wang, Chen Zhu, Jonas Geiping, Chaowei Xiao, and Tom Goldstein. 2023. On the Exploitability of Instruction Tuning. Advances in Neural Information Processing Systems (NeurIPS), 36:61836-61856.  
+Aviv Slobodkin, Eran Hirsch, Arie Cattan, Tal Schuster, and Ido Dagan. 2024. Attribute First, then Generate: Locally-attributable Grounded Text Generation. In Annual Meeting of the Association for Computational Linguistics (ACL), pages 3309-3344.
+
+Johanne R Trippas, Sara Fahad Dawood Al Lawati, Joel Mackenzie, and Luke Gallagher. 2024. What do Users really Ask Large Language Models? An initial Log Analysis of Google BARD Interactions in the Wild. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR), pages 2703-2707.  
+Ming Tu, Kevin Huang, Guangtao Wang, Jing Huang, Xiaodong He, and Bowen Zhou. 2020. Select, Answer and Explain: Interpretable Multi-Hop Reading Comprehension over Multiple Documents. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), volume 34, pages 9073-9080.  
+Vijay Viswanathan, Kiril Gashteovski, Kiril Gashteovski, Carolin Lawrence, Tongshuang Wu, and Graham Neubig. 2024. Large language models enable few-shot clustering. Transactions of the Association for Computational Linguistics (TACL), 12:321-333.  
+Denny Vrandecic and Markus Krötzsch. 2014. WikiData: A Free Collaborative Knowledge Base. Communications of the ACM, 57(10):78-85.  
+Liang Wang, Nan Yang, Xiaolong Huang, Binxing Jiao, Linjun Yang, Daxin Jiang, Rangan Majumder, and Furu Wei. 2022. Text Embeddings by Weakly-Supervised Contrastive Pre-training. arXiv preprint arXiv:2212.03533.  
+Zhao Xu, Wiem Ben Rim, Kiril Gashteovski, Timo Sztyler, and Carolin Lawrence. 2024. A Human-Centric Evaluation Platform for Explanable Knowledge Graph Completion. In European Chapter of the Association for Computational Linguistics (EACL): System Demonstrations, pages 18-26.  
+Yi Yang, Wen-tau Yih, and Christopher Meek. 2015. WikiQA: A Challenge Dataset for Open-Domain Question Answering. In Empirical Methods in Natural Language Processing (EMNLP), pages 2013-2018.  
+Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W. Cohen, Ruslan Salakhutdinov, and Christopher D. Manning. 2018. HotpotQA: A Dataset for Diverse, Explainable Multi-hop Question Answering. In Conference on Empirical Methods in Natural Language Processing (EMNLP).  
+Zhangyue Yin, Yuxin Wang, Xiannian Hu, Yiguang Wu, Hang Yan, Xinyu Zhang, Zhao Cao, Xuanjing Huang, and Xipeng Qiu. 2023. Rethinking Label Smoothing on Multi-hop Question Answering. In China National Conference on Chinese Computational Linguistics, pages 72-87. Springer.  
+Xiang Yue, Boshi Wang, Kai Zhang, Ziru Chen, Yu Su, and Huan Sun. 2023. Automatic Evaluation of Attribution by Large Language Models. In Findings of ACL: Conference on Empirical Methods in Natural Language Processing (EMNLP Findings), pages 4615-4635.
+
+Jiahao Zhang, Haiyang Zhang, Dongmei Zhang, Liu Yong, and Shen Huang. 2024. End-to-End Beam Retrieval for Multi-Hop Question Answering. In North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT), pages 1718-1731.  
+Shengyu Zhang, Linfeng Dong, Xiaoya Li, Sen Zhang, Xiaofei Sun, Shuhe Wang, Jiwei Li, Runyi Hu, Tianwei Zhang, Fei Wu, et al. 2023. Instruction Tuning for Large Language Models: A survey. arXiv preprint arXiv:2308.10792.  
+Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. 2020. BERTScore: Evaluating Text Generation with BERT. In Conference on Learning Representations (ICLR).
+
+# A Comprehensive Discussion on Related Work
+
+We split the related work papers into several categories: tasks, datasets, methods and metrics.
+
+# A.1 Context Attribution Tasks
+
+Attributable to Identified Sources (AIS): given a generative text  $t_{g}$  and a context text  $t_{c}$ , is  $t_{g}$  attributable to  $t_{c}$ ? Rashkin et al. (2023) propose a manual framework that defines the AIS task and evaluates the AIS scores across several NLP tasks, namely conversational QA (Anantha et al., 2020; Dinan et al., 2019), text summarization (Nallapati et al., 2016) and table-to-text (Parikh et al., 2020). Our work differentiates in three important aspects: (1) we focus on a broader QA setup (i.e., single-question QA and conversational QA), which makes our work a subset of the broader AIS task; (2) we focus on more fine-grained level: our attributions are not on the entire text level, but rather on a sentence level, which has been shown in user studies to be more useful to end-users (Slobodkin et al., 2024); (3) the AIS task entails a manual evaluation framework, while our work provides automatic evaluation with golden data.
+
+In-line Citation Generation: uses LLMs to produce in-line citations along with the generated text (Bohnet et al., 2022; Gao et al., 2023b; Huang et al., 2024b). This typically works on paragraph or document level. Slobodkin et al. (2024) propose a fine-grained task, where the attributions are on sentence level, because such granularity is more useful to human end users. Because generating such in-line citations can result in producing completely made up citations, Yue et al. (2023) propose a task that checks whether the in-line generated citations from LLMs are actually attributable or not. Instead of using binary attributable/non-attributable labels (like with AIS), they propose more fine-grained labels for this problem: attributable, extrapolatory, contradictory and non-attributable. Contrary to such approaches, our work focuses on post-hoc context attributions: given an answer to a question, find the sentences in the context that support the factuality of the answer.
+
+Post-hoc Attribution: determines which parts of the context are attributable to an already answered question (Yang et al., 2018). Within the post-hoc context, there are two other subcategorizations of the task: contributive and corroborative post-hoc
+
+attribution (Cohen-Wang et al., 2024).
+
+Post-hoc Attribution (Contributive): ContextCite (Cohen-Wang et al., 2024) and Mirage (Qi et al., 2024) define a post-hoc task that aims at discovering which parts of the context caused the LLM to generate the particular response. Their evaluation methods, however, are based on proxy metrics that do not rely on golden annotations, while in our work we rely on automatic annotations that rely on golden data.
+
+Post-hoc Attribution (Corroborative): this task is similar to contributive post-hoc attribution. The difference is that the constraint for causality is not necessarily enforced, but should support the factuality of the statement (Cohen-Wang et al., 2024). Many works are based on coarse-grained level and provide attributions on either paragraph level (Menick et al., 2022), document level (Nakano et al., 2021) or on multi-document level, where they have a RAG component that retrieves the documents that are potentially attributable (Gao et al., 2023a; Buchmann et al., 2024).
+
+Context Attributions to other Modalities. Other line of work maps the attributions to other modalities, such as knowledge graphs Dammu et al. (2024). Similarly, Maheshwari et al. (2024) take multi document collection as input, construct a graph of narratives, and then generate a presentation (i.e., slides) for the topic, along with attributions from the generated content of the slides with the original documents. We do not investigate such cases, and focus on attributing answers to sentences within the user-provided context.
+
+Post-hoc Attribution for Text Summarization. Ernst et al. (2021) proposed a task, dataset and baseline model (dubbed SuperPAL) for detecting attributions for text summarization. In a followup work, Ernst et al. (2022) an extension of the task, this time for clustering propositions for text summarization and Ernst et al. (2024) extend this to multi-document summarization. Krishna et al. (2023) investigate whether such text summarization alignments are helpful for humans. In our work, we focus on the question answering (QA) task.
+
+# A.2 Datasets
+
+AIS. Rashkin et al. (2023) proposed the AIS dataset, which contains three tasks: question answering, table-to-text and text summarization. Here, for each data point, there is a query and an
+
+LLM-generated response, along with label by humans whether it is fully attributable or not. This data is on paragraph and document level, and lacks the granularity of a sentence level. Therefore, we do not use it in our work.
+
+HotpotQA. With HotpotQA (Yang et al., 2018), the authors propose an explainable multi-hop QA dataset. The dataset also contains attribution links (i.e., explanations) for the answers: spans of text that belong to the input context, which are supporting the statement in the answer. The authors set up baselines for measuring the ability of attributions of models on sentence level, which is in line with what we do. In our work, we integrated HotpotQA as part of our setup for both training and testing.
+
+AttributionBench. This is a benchmark for attribution evaluation of LLM generated content (Li et al., 2024). In particular, the benchmark assesses whether the assigned attribution on a generated text is actually attributable. In particular, given a query, response set  $\mathcal{R}$  (containing claims) and evidence set  $E$ , the task is to label as "attributable" or "not attributable" every claim against  $E$ . This work operates on a coarse-grained level (paragraphs or whole documents). Similarly, Yue et al. (2023) proposed another dataset for evaluating attribution of LLM-generated text, same on paragraph level. In our work, we focus on sentence-level context attribution. Therefore, we did not include this dataset in our work.
+
+Conversational QA. CoQA (Reddy et al., 2018) is a conversational QA dataset, which contains a context, questions-answer pairs between two people (teacher and student) and sentence-level supporting evidence for the context. We use this dataset in our evaluation to test the out-of-domain capabilities of LLMs for context attribution. We also use QuAC (Choi et al., 2018) and ORConvQA (Qu et al., 2020), which are conversational QA datasets similar to CoQA.
+
+QASPER. This dataset is from the scientific domain (Dasigi et al., 2021). The dataset contains title and an abstract of a paper, question and answer about the content. This data has information on paragraph level, not on sentence level. Therefore, we do not use it in our experiments.
+
+WikiQA. Dammu et al. (2024) use WikiQA (Yang et al., 2015), because it's Wikipedia-based dataset, which can be linked to Wikipedia-derived
+
+KG like Wikidata (Vrandecic and Krötzsch, 2014). In our work, we focus only on text modality, which is why we do not include this dataset into our evaluation.
+
+WikiNLP. The WikiNLP dataset (Gashteovski et al., 2019) is a dataset that contains the entire English Wikipedia, along with linguistic annotations (e.g., POS tags, dependency parse trees, etc.) and semantic annotations (e.g., NER tags and entity links). We use this dataset for the synthetic data generation, because it keeps the linked information from the Wikipedia articles, which are annotated by humans. For this reason, the dataset has been used in wide range of tasks in research, mostly for information extraction (Dukić et al., 2024; Kotnis et al., 2023, 2022b; Gashteovski et al., 2020), but also for other tasks such as clustering (Viswanathan et al., 2024), open link prediction (Broscheit et al., 2020) and entity linking (Nanni et al., 2019; Radevski et al., 2023)
+
+# A.3 Metrics and Evaluation
+
+AIS. The AIS framework (Rashkin et al., 2023) is human annotation framework. Given a generated text chunk and a context chunk (this can be sentence, paragraph or document), a human evaluated whether the generated text chunk is fully attributable or not. It is basically a binary classification problem. In their data, the authors focus on document level granularity, which is not useful for humans. In our setup, we check for each sentence in the context if it supports the answer.
+
+AutoAIS. To evaluate the attributed information, Slobodkin et al. (2024) use AutoAIS metric: an NLI-based scoring. Prior studies have shown that this metric highly correlated with human annotations (Bohnet et al., 2022; Gao et al., 2023b). This is an extension to the AIS metric. We do not use proxy metrics, but rely on golden annotations by humans.
+
+AttrScore. With AttrScore, Yue et al. (2023) consider LLM generated content on the one hand and citation documents on the other hand. Then, the score evaluates whether a provided citation is attributable, extrapolatory, contradictory or non-attributable. Essentially, it is an extension of AIS, such that it provides more fine grained labels for the provided citations. The AttrScore is basically a fine-tuned LLM that provides these scores.
+
+Unsupervised Metrics. ContextCite proposed the Top-k-drop and LDS metric to evaluate the causal post-hoc attribution. These metrics do not require labeled data. Berchansky et al. (2024) uses ROUGE and BERTScore to evaluate their results. We do not use unsupervised metrics and rely on automated evaluation with golden annotations by humans.
+
+# A.4 Methods
+
+Multihop QA. There has been significant amount of work on tackling the context attribution problem on sentence level for multi-hop QA (Zhang et al., 2024; Ho et al., 2023; Yin et al., 2023; Fu et al., 2021; Tu et al., 2020; Fang et al., 2020). While we also investigate this problem, in contrast to our work, these works focus only on the multihop QA task. In our work, we also explore other QA setups, including conversational QA with different domains. Moreover, these papers do not investigate the capabilities of LLMs regarding the context attribution problem, but rather propose specific methods that are tailor-made for the multihop QA problem, which involves both answering the questions and providing supporting sentences to the answers.
+
+In-line Citation Generation. Another line of work focuses on guiding LLMs to generate in-line citations along with the generated text (Li et al., 2023). Slobodkin et al. (2024) tackle this problem on a sentence level, but do not investigate the post-hoc context attribution case. Moreover, they rely on proxy metrics such as AutoAIS (Gao et al., 2023a) and BERTScore (Zhang et al., 2020). Similarly, Bohnet et al. (2022) proposes methods for in-line citation generation, but this work is more coarse-grained and focuses on paragraph and document level. They also report their findings on proxy metrics. Their method is based on retrieval and they do not investigate the LLMs' capabilities thoroughly. Gao et al. (2023b) assign citations to LLM-generated content, where they retrieve the information from a large collection of documents (also, it's on paragraph and document level, not on sentence level). START (Huang et al., 2024b) propose a data synthetic generation method for in-line citation generation on document level, where each citation refers to an entire document. FRONT (Huang et al., 2024a) also investigates synthetic data generation of in-line citation generation, where the citations assigned to the sentences in the
+
+output are entire documents. Similarly, Schimanski et al. (2024) propose a synthetic data generation pipeline for fine-tuning models that solve the same problem. Berchansky et al. (2024) use Chain-of-Thought approaches and fine-tuning smaller LLMs in order to solve this problem. Patel et al. (2024) also fine-tune a model specifically for this task, and the attributions are on paragraph level.
+
+Post-hoc Context Attribution. Ramu et al. (2024) propose template-based in-context learning method for post-hoc context attribution. In particular, they use standard retrievers as a first step to pre-rank the text (e.g., BM25 and dual encoders (Ni et al., 2022)) and then they use LLMs to classify (i.e., rerank) the relevant sentences. ContextCite (Cohen-Wang et al., 2024) uses ablation-based methods to infer the attributions of post-hoc generated text.
+
+User Study. Recent work has called for a more human-centric research in NLP (Kotnis et al., 2022a). In such work, the idea is to involve the user (i.e., the final stakeholder) in the process of research, which is typically done with some forms of user studies (Rim et al., 2024; Xu et al., 2024; Ilievski et al., 2024). In this spirit, we want to verify whether our models are useful for end users. To this end, we performed a user study, whereas users were asked to solve a fact-checking task with the use of our context attribution model. We found that our approach does indeed make human end-users faster in performing the manual fact-checking task (for details of our user study, see Section 3.3.5).
+
+# B Method for Synthetic Data Generation
+
+# B.1 Multi-hop Generation of Attribution Data
+
+To generate synthetic data with the use of Wikipedia, we use the WikiNLP dataset (Gashteovski et al., 2019). It contains the text from all Wikipedia articles along with annotations for links within the text that link to other Wikipedia articles. The main idea is to use the links in order to imitate reasoning hops across different (related) articles. Therefore, we filter out all articles that either do not contain links or that contain links to articles that do not contain links. Finally, for each article, we use only the first paragraph, because this is considered to be the paragraph that contains the most "definitional information" (Bovi et al., 2015); i.e., information that precisely describes the target concept of the article and contains the most important
+
+information about it.
+
+Then, for each article, we randomly select a sentence that contains at least one link to another Wikipedia article. Each of the sentences that we sample serves as ground truth for the context attribution. With these sentences, we then prompt an LLM to generate a question-answer pair.
+
+# B.2 Question-Answer Pairs Generation
+
+To generate context attribution training data that requires multi-hop reasoning, using the multi-hop chain of sentences, we prompt the LLM (see §D.1 for the prompt we use) by providing it only the formed chain as the ground truth attributions, which the LLM must use for the question-answer pair generation. The LLM generates a question-answer pair that can only be answered using the information in these supporting sentences, ensuring the pairs are grounded in the provided evidence.
+
+For dialogue-centric context attribution data, we simply provide the LLM with a single Wikipedia article, and prompt the model to generate multi-turn QA pairs and provide the attribution sentence for each QA pair (see §D.2 for the prompt we use).
+
+# B.3 Distractors Mining
+
+In realistic scenarios, whether the context is user-provided or retrieved through RAG, the system typically encounters multiple context documents that are highly similar to those containing the evidence sentences. To bridge this gap between our synthetically generated training data using SYNQA and the data models encounter "in the wild", we augment each training sample with hard negative distractor articles. We obtain embeddings using E5 (Wang et al., 2022) for each Wikipedia article in our collection. Then, for each article containing a supporting sentence for the question-answer pair, we randomly sample up to three distractor articles that share semantic similarity with the ground truth article. This process increases the difficulty of the training data, producing models better equipped to handle diverse testing scenarios.
+
+# B.4 Comparison to HotpotQA
+
+Although our method is inspired by HotpotQA (Yang et al., 2018), note that we do not aim to recreate the HotpotQA dataset. Our method has significant differences, enabling us to generate as
+
+much data as needed, with a higher domain variability.
+
+Particularly, the method of Yang et al. (2018) is more curated and has a human in the loop in multiple steps. First, the authors manually select the target entities (and, with that, the target articles from which the annotators create the question and answer pairs). The reason for this is that many highly specialized articles—e.g., the article for IPv4 protocol—are not suitable for crowd-workers to both identify meaningful questions and provide answers for those questions. Our approach does not have this constraint and, therefore, produces data that has much higher domain variability.
+
+Second, their method uses Amazon Mechanical Turk workers to annotate the questions, answers, and attribution sentences. In our case, we automatically select the hopped sentences (which serve as gold context attribution data), and then we use these sentences to generate question-answer pairs with an LLM.
+
+Third, while HotpotQA always enforces multihop QA pairs, we do not instruct the LLM to do that. Rather, we first allow the LLM to decide whether generating such a multihop QA pair is possible for the incoming context attribution sentences. If so, then the LLM generates multihop QA pairs. Otherwise, it generates direct QA pairs that do not need hops; i.e., QA pairs like in SQuAD (Rajpurkar et al., 2016).
+
+Fourth, the HotpotQA annotation method does not allow for dialogue QA. In our method, we also create dialogue context attribution data.
+
+With these differences in mind, we showed that, compared to HotpotQA, our data generation method exhibits the following advantages: (1) generates data with higher domain variability; (2) goes beyond multi-hop QA and also generates direct QA pairs (like SQuAD) as well as dialogue QA data; (3) the data is generated in completely automatic manner without the involvement of humans.
+
+# C Multi-turn to single-turn conversations
+
+We convert the multi-turn Quac and CoQA datasets to single-turn (Quac-ST and CoQA-ST) using Llama 70B. In Table 4, we provide examples of the questions and answers before and after converting them.
+
+<table><tr><td>ID</td><td>Type</td><td>Original</td><td>Rephrased</td></tr><tr><td>Q1</td><td>Question</td><td>What happened in 1983?</td><td>What significant event occurred in Anna Vissi&#x27;s life in 1983?</td></tr><tr><td>A1</td><td>Answer</td><td>In May 1983, she married Nikos Karvelas.</td><td>In May 1983, Anna Vissi married Nikos Karvelas, a composer.</td></tr><tr><td>Q2</td><td>Question</td><td>Did they have any children?</td><td>Does Anna Vissi have any children?</td></tr><tr><td>A2</td><td>Answer</td><td>She gave birth to her daughter Sofia.</td><td>Anna Vissi gave birth to her daughter Sofia in November.</td></tr><tr><td>Q3</td><td>Question</td><td>What collaborations did she do with Nikos?</td><td>What is the nature of Anna Vissi&#x27;s collaborations with Nikos Karvelas?</td></tr><tr><td>A3</td><td>Answer</td><td>After their marriage, she started a close collaboration with Karvelas.</td><td>Since 1975, Nikos Karvelas has composed songs for all of Anna Vissi&#x27;s releases, which have become gold or platinum.</td></tr></table>
+
+Table 4: Examples of the original (Quac and CoQA) and rephrased questions and answers (Quac-ST and CoQA-ST).
+
+# D Prompts to generate SYNQA synthetic training data
+
+# D.1 Multi-hop reasoning context attributon
+
+# SYSTEMPROMPT
+
+You are tasked with generating a concise and focused question-answer pair using information from provided Wikipedia sentences. Follow these instructions carefully:
+
+1. You will be provided with multiple Wikipedia articles, each containing:  
+- The title of the article.  
+- One specific sentence from the article.  
+2. Your goal is to generate a short, factual question and a concise answer, ensuring:  
+- The question-answer pair is grounded in the provided sentences.  
+- The reasoning is logical, clear, and references all sentences used.
+
+# 3. Key Constraints:
+
+- Questions must address a single coherent topic or concept that can be logically inferred from the provided sentences.  
+- Avoid combining unrelated pieces of information into a single question.  
+- The "reasoning" must explain how each sentence in the "ids" field contributes to answering the question but should remain brief and to the point.
+
+4. Aim for brevity:  
+- Questions should be concise and avoid unnecessary details.
+
+- Answers should be short, typically no more than one sentence.  
+- Keep the reasoning concise, focusing only on the necessary logical connections.  
+5. Multi-hop reasoning is encouraged but must be natural and focused:  
+- Combine information only when it is logical and directly relevant to the question.  
+- Do not create overly complex questions that combine weakly related information.  
+6. Provide your response in raw JSON format with the following keys:  
+- "question": A concise and clear question string.  
+- "answer": A short and factual answer string.  
+- "ids": A list of JSON-compatible arrays (e.g., [[0, 0], [1, 0]]) representing the indices of all sentences used to generate the question-answer pair.  
+- "reasoning": A brief explanation of how each sentence in "ids" was used to generate the question-answer pair.
+
+# Important Notes:
+
+- Ensure the question-answer pair is entirely self-contained and logically consistent.  
+- Do not include unnecessary or weakly related information in the question or answer.  
+- Avoid introducing information not present in the provided sentences.  
+- Do not include additional formatting, explanations, or markdown in your response.
+
+# USER PROMPT
+
+Here are the titles and sentences:
+
+Title: [First Article Title]
+
+[0, 0] [First sentence from the article]
+
+Title: [Second Article Title]
+
+[1, 0] [Second sentence from the article]  
+Title: [Third Article Title]  
+[2, 0] [Third sentence from the article]
+
+Use the provided sentences to generate a question-answer pair following the specified guidelines. Respond only in raw JSON with no additional formatting or markdown.
+
+Given the SYSTEM and the USER prompt, the LLM is generating the question-answer pair, which when combined with the full articles, yields a single SYNQA training data sample.
+
+# D.2 Dialogue context attributon
+
+# SYSTEMPROMPT
+
+You are an AI assistant that generates structured question-answer pairs based on a passage. Your goal is to create meaningful, factual, and reasoning-based questions that require connecting multiple sentences.
+
+Follow these strict guidelines:
+
+- Format the output as a valid JSON array, where each item has:  
+- "question": A clear, concise question.  
+- "answer": A short, factual response.  
+- "sentence-numbers": A list of integers pointing to all relevant supporting sentences.  
+- Ensure questions are generated in a random sentence order (not sequential).  
+- Some questions must reference multiple sentences for reasoning.  
+- Some sentences should be reused across multiple questions.  
+- Later questions should rely on earlier information and use pronouns or indirect references to maintain logical flow.  
+- Introduce a mix of fact-based, causal, and inference questions.  
+- Avoid introducing information not present in the passage.  
+- Ensure all relevant sentences are cited for each answer.
+
+Your response must be valid JSON containing 5 to 10 question-answer pairs.
+
+and the user prompt:
+
+# USER PROMPT
+
+Here is a passage:
+
+Title: [Title of the passage]
+
+0. [First sentence of the passage]  
+1. [Second sentence of the passage]  
+2. [Third sentence of the passage]  
+3. [Fourth sentence of the passage]
+
+···
+
+Generate structured question-answer pairs following these constraints:
+
+- Return output in JSON format only: ["question": "...", "answer": "...", "sentenceNumbers": [...], ...]  
+- Use random sentence order, not sequential.  
+- Some questions should require multiple sentences.  
+- Some sentences should be reused across different Q&A pairs.  
+- Later questions must reference earlier ones using pronouns or indirect mentions.  
+- Include a mix of question types:  
+- Factual questions that can be answered directly from the passage.  
+- Causal questions that require understanding relationships between sentences.  
+- Inference-based questions that require implicit reasoning.  
+- Ensure sentence numbers fully cover the reasoning required.
+
+Return only JSON, with no extra text.
+
+# E Prompts for zero-shot models
+
+In order to obtain context attributions with the instruction-tuned LLM (Jiang et al., 2023; Dubey et al., 2024), we use the following prompts:
+
+# SYSTEMPROMPT
+
+You are an AI assistant that identifies the sentence(s) in a provided context document most relevant for answering a specific question. Your task is to select only the sentence(s) containing the explicit information needed to answer the question accurately, without adding extra context.
+
+# USER PROMPT
+
+Context Document:
+
+[numbered sentences from the context]
+
+Question: [query text]
+
+Answer: [answer text]
+
+Based on the context document, identify the sentence number(s) from the following choices: [list of numbers]. Select only the sentence(s) that contain explicit information needed to answer the question directly. Answer only with the corresponding number(s) in parentheses, without additional explanation.
+
+# F User Study
+
+An example of the attribution scenario evaluated in our user study. See Figure 5 for details.
+
+![](images/083ee0306508bf42ab0c44a40408fb9d56deedeb488ebb0381c2e93ca8c787bc.jpg)  
+Figure 5: An example of the attribution scenario evaluated in our user study. Both the answer and the context attributions are highlighted to help the user verify the correctness of the answer. In the absence of highlights, the user is instructed to read the entire context. This example showcases a practical application of context attribution in real-world interactions with LLM-generated content.

ACL/2025/On Synthesizing Data for Context Attribution in Question Answering/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7960824537acfbce56b8715ba0cd4cc3cc4151af70f18fa1ab4deb55983d66c3
+size 642260

ACL/2025/On Synthesizing Data for Context Attribution in Question Answering/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:60dfe25ba3e9b50a64522ae718d984785e211b362f166ae8a201b094ca928039
+size 611827

ACL/2025/On Synthetic Data Strategies for Domain-Specific Generative Retrieval/82609c3a-8fd4-4dd1-873d-a6c74cf6fdf5_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8c1905ecc5e1b32b0eabebde86f6210832d827fb09ddc3607b26b72969e846aa
+size 111758

ACL/2025/On Synthetic Data Strategies for Domain-Specific Generative Retrieval/82609c3a-8fd4-4dd1-873d-a6c74cf6fdf5_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3e9a2c020152bf3b1e0bea30c09d020fff3e2079902e9e76851cf0af5d68678b
+size 133917

ACL/2025/On Synthetic Data Strategies for Domain-Specific Generative Retrieval/82609c3a-8fd4-4dd1-873d-a6c74cf6fdf5_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:924bb3a76a144723884a9705f366e6a25a699d970d9a83ea05c3d7bd120a5588
+size 612975

ACL/2025/On Synthetic Data Strategies for Domain-Specific Generative Retrieval/full.md

@@ -0,0 +1,465 @@
+# On Synthetic Data Strategies for Domain-Specific Generative Retrieval
+
+Haoyang Wen $^{\ddagger}$ , Jiang Guo $^{\dagger*}$ , Yi Zhang $^{\dagger}$ , Jiarong Jiang $^{\dagger}$ , Zhiguo Wang $^{\dagger}$
+
+$^{\ddagger}$ Language Technologies Institute, Carnegie Mellon University  $\dagger$ AWS AI
+
+hwen3@cs.cmu.edu
+
+{gujiang, imyi, jiarongj, zhiguow}@amazon.com
+
+# Abstract
+
+This paper investigates synthetic data generation strategies in developing generative retrieval models for domain-specific corpora, thereby addressing the scalability challenges inherent in manually annotating in-domain queries. We study the data strategies for a two-stage training framework: in the first stage, which focuses on learning to decode document identifiers from queries, we investigate LLM-generated queries across multiple granularity (e.g. chunks, sentences) and domain-relevant search constraints that can better capture nuanced relevancy signals. In the second stage, which aims to refine document ranking through preference learning, we explore the strategies for mining hard negatives based on the initial model's predictions. Experiments on public datasets over diverse domains demonstrate the effectiveness of our synthetic data generation and hard negative sampling approach.
+
+# 1 Introduction
+
+Generative retrieval is emerging as a promising paradigm for information retrieval (IR), leveraging generative models (e.g., Transformers, Vaswani et al., 2017) to directly produce ranked lists of potentially relevant document identifiers for a user query. Although prior work has made progress on various fronts, including training strategies (e.g., identifier choices) (Tay et al., 2022; Zhou et al., 2022; Bevilacqua et al., 2022; Sun et al., 2023), modeling techniques (Chen et al., 2023b; Zhou et al., 2023; Li et al., 2024b), and inference methods (Cao et al., 2020; Lee et al., 2023a; Zhang et al., 2024), the role of data strategies in training generative retrieval models, particularly when dealing with domain-specific corpora, remains relatively underexplored. This gap is critical: as generative retrieval models internalize entire corpus within their parametric memory, the choice and quality of
+
+training data are likely to play a critical role in their performance.
+
+To mitigate the high cost and scalability challenges of in-domain annotation, most studies have adopted DSI-QG (Zhuang et al., 2022a), which uses passage-level synthetic queries generated by docT5query (Nogueira et al., 2019) (a model trained on MS-MARCO data). However, applying such off-the-shelf synthetic data strategies to new domains may not suffice. Unlike dense retrieval approaches, which focuses on strong text representation (Karpukhin et al., 2020; Izacard et al., 2021), generative retriever must develop three key capabilities: (1) memorization (storing the content of the corpus (e.g., documents) and mapping them to their assigned identifiers), (2) generalization (inferring beyond explicit textual cues from user queries), and (3) relevance scoring (accurately ranking document identifiers by relevance to a given query). Domain-specific corpora can amplify these challenges, as the model must adapt its internal representations to reflect domain nuances while maintaining robust generalization and ranking accuracy. In this work, we systematically investigate data strategies that can foster these core capabilities.
+
+We introduce a two-stage training framework. The first stage focuses on mapping an input directly to document identifiers via supervised fine-tuning on synthetic data. The second stage uses preference learning to further enhance the ranking performance (Zhou et al., 2023; Li et al., 2024b). Here we adopt Regularized Preference Optimization (Pang et al., 2024, RPO), an effective alternative to PPO-based reinforcement learning (Ouyang et al., 2022). We study the data strategies for both stages.
+
+The first stage focuses on the memorization and generalization ability. We examine two data sources as the input for decoding document identifiers during training: the context data (e.g., chunks) directly extracted from the corpus, and synthetic queries that represent various relevance signals.
+
+For synthetic queries, we investigate query generation using multi-granular context (e.g., sentence-level, chunk-level) to capture both local and global information from the corpus. We also explore adding constraints derived from available metadata or domain-specific knowledge when generating synthetic queries to enhance the model's ability of handling complex, domain-relevant queries.
+
+In the first stage, models are optimized to produce a single positive candidate, without considering relevance between different candidates. In the second stage, we further create data to enhance the model's ranking capability through preference learning (Zhou et al., 2023; Li et al., 2024b). We study the selection of negative candidate documents for preference learning. Rather than relying on static offline data, we collect preference data online from the model's own top-ranked candidates after the first stage, and compare it to random sampling from the corpus. We further investigate the choices and impact of varying the number of negative candidates to the ranking performance.
+
+We conduct experiments on datasets covering various aspects of relevance, including the widely adopted Natural Questions (Kwiatkowski et al., 2019, NQ), a multi-hop dataset MultiHopRAG (Tang and Yang, 2024), and two perspective-based retrieval datasets: AllSides (Baly et al., 2020) and AGNews (Yu et al., 2023) from Zhao et al. (2024). We show that queries with different aspects, such as multi-granular and constrains-based queries, significantly improve the retrieval performance compared to relying solely on chunk-level synthetic queries from query generation models. Additionally, upsampling context data further improves the performance. Moreover, we show that these data strategies generalize well to other types of document identifiers, such as atomic identifiers. Finally, we demonstrate that RPO effectively improves the ranking performance of generative retrieval, and the key lies in the selection of high-quality negative candidates: high-quality hard negative candidates improve the performance while random negatives may have an adverse impact.
+
+In summary, this work offers a comprehensive investigation of data strategies for building scalable and effective domain-specific generative retrieval systems. Our findings emphasizes the importance of creating high-quality and diverse synthetic queries that capture multiple levels of granularity within the corpus, as well as informed negative selection strategies for ranking optimization.
+
+# 2 Generative Retrieval Framework
+
+A typical generative retrieval framework takes a query as input, and generates the corresponding relevant document identifiers as the retrieval results (Tay et al., 2022). Because each document in the corpus has a unique identifier, one can then use these identifiers to retrieve the corresponding documents for downstream tasks.
+
+# 2.1 Document Identifiers
+
+We primarily use semantic document identifiers in our experiments due to their superior performance and better scalability to larger corpora. Instead of using corpus-specific semantic identifiers like titles or URLs, we adopt a more general, keyword-based approach that can be applied to a wide variety of corpora (Zhou et al., 2022). Specifically, we instruct an LLM to produce a list of keywords that describes the content of a document, and use this keyword list as its semantic identifier.
+
+In addition, we extended our synthetic data strategies to other types of identifiers to validate its generalizability, such as atomic identifiers (Tay et al., 2022), which are unique tokens that can be generated through a one-step decoding or classification process.
+
+# 2.2 Generative Modeling
+
+The generative retrieval model learns to generate the identifier of a relevant document given a query. Formally, for a query  $q$  and a relevant document  $d$  with identifier  $d'$ , generative retrieval aims to produce  $d'$  given  $q$ , which can be represented as:
+
+$$
+\begin{array}{l} \operatorname {s c o r e} (q, d) = P \left(d ^ {\prime} \mid q; \theta\right) \\ = \prod_ {i} P \left(d _ {i} ^ {\prime} \mid d _ {<   i} ^ {\prime}, q; \theta\right), \\ \end{array}
+$$
+
+where  $d_i^l$  is the  $i^{\mathrm{th}}$  token of the identifier. To ensure the generated identifiers are valid during inference, we use constrained beam search with Trie (Cao et al., 2020) to restrict the output token space at each decoding step. The top- $k$  output from the beam search serves as the final retrieval results.
+
+Compared to dense retrieval models (Karpukhin et al., 2020), generative retrieval bypasses the need for an external index by directly producing relevant document identifiers. However, there are distinct challenges in learning a generative retrieval model. As it solely relies on parametric knowledge, the model must not only learn the retrieval task, but also capture and encode document content in a way
+
+![](images/b6ce90d897349449973731ba8a58f78656a5943af1486d9931046703979a9c24.jpg)  
+Figure 1: The overall workflow of the generative retrieval training and synthetic data utilization at each stage.
+
+that associates each document with its identifier. Therefore, generative retrieval often requires training on the entire corpus to enable the model to memorize and comprehend the necessary information effectively.
+
+# 3 Supervised Fine-Tuning Data Strategy
+
+In a typical domain-specific setup, we often assume access to a corpus with limited or no labeled data for domain-specific training (Hashemi et al., 2023). Therefore, it is crucial to create high-quality synthetic data that thoroughly covers the entire corpus for generative retrieval training.
+
+Our synthetic data comprises two main components: Context2ID data and Query2ID data. Context2ID involves training the model to retrieve the document identifiers given the document's content. Query2ID focuses on teaching the model to retrieve relevant document identifiers from a given query. Combining these two data types encourages the model to learn both content memorization and retrieval given a query.
+
+# 3.1 Supervised Fine-Tuning Objective
+
+At this stage, we train the model to generate relevant document identifiers by maximizing the probability of each individual token. While typical supervised fine-tuning (SFT), especially with encoder-decoder architectures such as T5, focuses on optimizing the output sequence (i.e. document identifiers), it's also part of the training goal for generative retrieval models to comprehend and memorize the context. To this end, we also optimize the model for learning to decode the input. Specifically, for a given query-document pair  $(q,d)$ , where  $q$  could be an actual query or a text chunk from the document, the model maximizes the likelihood of
+
+the combined input and output sequence:
+
+$$
+\begin{array}{l} \mathcal {L} _ {\mathrm {s f t}} (q, d) = - \log P (d ^ {\prime}, q; \theta) \\ = - \sum_ {i} \log P (q _ {i} \mid q _ {<   i}; \theta) \\ - \sum_ {i} \log P \left(d _ {i} ^ {\prime} \mid d _ {<   i} ^ {\prime}, q; \theta\right). \\ \end{array}
+$$
+
+# 3.2 Context2ID
+
+Context2ID data is created by pairing each chunk of text in the corpus with its corresponding document identifier. The goal of Context2ID data is to help the generative retrieval model associate each document's content with its unique identifier, i.e., "memorizing" the text.
+
+# 3.3 Query2ID
+
+Query2ID is designed to teach the model to retrieve the relevant document identifiers given a query. It helps the model to learn the core retrieval task and also further comprehend content from the query perspective.
+
+Previous work (Zhuang et al., 2022a) finds it effective to use a query generation model (e.g., docT5query, Nogueira et al., 2019) to produce synthetic queries from documents using multiple independent samplings. In this work, we instead use an LLM for synthetic query generation. Specifically, given a context (e.g., a document chunk), the LLM is instructed to generate a diverse set of  $m$  queries, thereby covering a wider range of semantic variations compared to the sampling-based approach with a specialized query generation model.
+
+# 3.3.1 Multi-Granular Query Generation
+
+We first generate queries with context at different levels of granularity: chunk-level and sentence-level. Chunk-level synthetic queries are produced
+
+<table><tr><td>Data Type</td><td>Example</td></tr><tr><td>Context</td><td>title: Christmas Day preview: 49ers , Ravens square off in potential Super Bowl sneak peek... source: Yardbarker ... San Francisco has racked up an NFL-leading 25 turnovers and has given up the second-fewest rushing yards (1,252) , . . .</td></tr><tr><td>Chunk-level query</td><td>What is the potential implication of this matchup between the 49ers and Ravens ?</td></tr><tr><td>Sentence-level query</td><td>Where does the 49ers ’ defense stand in terms of total yards allowed per game?</td></tr><tr><td>Constraints-based query</td><td>According to the Yardbarker article, which team has the league’s most effective running game?</td></tr></table>
+
+Table 1: Examples of different synthetic queries generated from MultiHop-RAG corpus.
+
+by providing the entire chunk as input to the LLM to capture higher-level semantics or facts, while sentence-level synthetic queries are produced by only providing individual sentences to focus on more specific details within the document Contractely, for each chunk, we ask the LLM to produce  $m_{c}$  chunk-level queries. We then split the chunk into individual sentences and ask the LLM to generate  $m_{s}$  sentence-level queries for each sentence.
+
+# 3.3.2 Constraints-Based Query Generation
+
+A key advantage of using an LLM for query generation is its ability to incorporate domain-specific instructions. For instance, we can prompt the LLM to include metadata constraints, such as the author name or political polarity of a document, in the generated queries. Although the specific constraint types depends on the metadata available and can be domain or dataset specific, they are common in real-world scenarios such as enterprise data. Table 10 in Appendix specifies the attributes that we use to produce constraints-based synthetic queries for each dataset. We ask the LLM to generate  $m_{i}$  queries for each document that incorporate these constraints, allowing our generative retrieval model to handle more specialized or domain-specific queries.
+
+# 4 Preference Learning Data Strategy
+
+Previous work (Zhou et al., 2023; Li et al., 2024b) have shown that incorporating ranking tasks can further enhance the relevance modeling of generative retrieval models. However, when generative retrieval models are based on large language models, complex ranking objectives – such as listwise optimization – often become computationally inefficient due to multiple forward passes. In this work, we instead use a simpler method and adopt the regularized preference optimization algorithm to perform the preference optimization, a technique widely applied in optimizing large language models. We will first briefly introduce the preference
+
+optimization method, and then turn our focus on the synthetic data construction, which consists of the synthetic queries along with their corresponding preferred or rejected candidates.
+
+# 4.1 Preference Optimization Objective
+
+We use Regularized Preference Optimization (Pang et al., 2024, RPO) as our optimization method for preference learning. It is an extended version of Directed Preference Optimization (Rafailov et al., 2023, DPO), including additional supervised fin-tuning loss to alleviate the over-optimization issues on negative responses. It takes an input query  $q$ , a positive candidate  $d_p$ , and a negative candidate  $d_n$  as input. The loss is in favor of the positive candidate while against the negative candidate
+
+$$
+\begin{array}{l} \mathcal {L} _ {\mathrm {r p o}} (q, d _ {p}, d _ {n}) = - \log \delta \left(\beta \log \frac {P \left(d _ {p} ^ {\prime} \mid q ; \theta\right)}{P \left(d _ {p} ^ {\prime} \mid q ; \theta_ {\mathrm {r e f}}\right)}\right) \\ \left. - \beta \log \frac {P \left(d _ {n} ^ {\prime} \mid q ; \theta\right)}{P \left(d _ {n} ^ {\prime} \mid q ; \theta_ {\text {r e f}}\right)}\right) \\ - \alpha \frac {\log P (d _ {p} ^ {\prime} \mid q ; \theta)}{\left| d _ {p} ^ {\prime} \right|}, \\ \end{array}
+$$
+
+where  $\theta_{\mathrm{ref}}$  is the parameter of the reference model, i.e., the supervised fine-tuned model from the first stage training.  $d_p^{\prime}$  and  $d_n^\prime$  are the identifiers of the positive and negative candidate, respectively.
+
+# 4.2 Synthetic Queries
+
+Similar to the previous section, in a domain-specific setup, we assume that we do not have enough data for model training. Therefore, after the supervised fine-tuning stage, we need a batch of new synthetic queries for preference learning.
+
+We still adopt the LLM-based query generation as with the supervised fine-tuning stage. However, there are a few key differences in the instructions. First of all, we ask the LLM to make queries as difficult as possible. At the same time, we ask the LLM to provide not only the synthetic queries but
+
+also their corresponding answers. This is to ensure that, while making difficult queries, those synthetic queries are still answerable using the given context.
+
+These changes make the new batch of synthetic queries different from queries used during supervised fine-tuning so that the model will not be over-optimized to the same batch of data. Intensifying the difficulties also increases the likelihood that the initial generative retrieval model makes mistakes, and therefore the model will benefit from the preference learning by learning from those mistakes.
+
+# 4.3 Candidate Selection
+
+After producing the synthetic queries, the next step is to select document candidate pairs for RPO optimization. For each training instance, we need one positive candidate and one negative candidate. As we always produce synthetic queries based on a document, the positive candidate can be naturally assigned. Therefore, the focus will be on selecting negative candidates for each synthetic query.
+
+To increase the hardness of the negative candidates, we choose to select negative candidates from the retrieval results. Specifically, after the supervised fine-tuning stage, we will use the generative retrieval model to perform retrieval on the synthetic queries for preference learning. Our strategy mainly focuses on selecting the top- $k$  negative candidates with ranks higher than the positive candidate from the retrieval results. In this way, if the positive candidate ranks in the top-1, we will not use the query for preference learning. If the rank of the positive candidate is higher than  $k$ , then there will be different numbers of negative candidates, depending on the rank. If the rank is lower than  $k$ , there will be  $k$  different negative candidates. When there are multiple negative candidates, we pair each negative candidate with the positive one to form a candidate pair instance for preference learning.
+
+# 5 Experiments
+
+# 5.1 Datasets
+
+We choose 4 datasets for our experiments: three domain-specific corpora - MultiHop-RAG (Tang and Yang, 2024), -AllSides (Baly et al., 2020) and AGNews (Yu et al., 2023) from Zhao et al. (2024) - as well as the general-domain dataset Natural Questions dataset (Kwiatkowski et al., 2019, NQ).
+
+For AllSides and AGNews, we mainly adopt queries from Zhao et al. (2024). In the case of AGNews, we replace the similar document part in
+
+<table><tr><td>Context Source</td><td>HIT@4</td><td>HIT@10</td><td>MAP@10</td><td>MRR@10</td></tr><tr><td>Chunk</td><td>43.64</td><td>66.65</td><td>13.98</td><td>31.14</td></tr><tr><td>Chunk+Sent</td><td>61.64</td><td>81.69</td><td>22.13</td><td>47.20</td></tr></table>
+
+Table 2: Ablation study on the effect of using synthetic queries generated with context at different levels of granularity on Multihop-RAG.
+
+queries with another attribute of perspective, as we focus on the query retrieval rather than document similarity search.
+
+For NQ, we use the "old document" split from Kishore et al. (2023), which constructs a subset of Wikipedia pages containing all positive candidates for training and testing, while keeping the corpus size manageable for generative retrieval training.
+
+# 5.2 Experiment Setup
+
+For all datasets, we use Mistral 7b (Jiang et al., 2023) series as the generative retrieval base model. We use Mixtral  $8\mathrm{x}7\mathrm{b}$  (Jiang et al., 2024) to generate all the synthetic queries and we use Claude 3 Sonnet (Anthropic, 2024) to generate keywords. Please refer to Appendix A for more details on the training infrastructure, hyperparameters, dataset-specific setup and statistics.
+
+# 5.3 Results
+
+We will discuss our experiment results for each of the stages. In the supervised fine-tuning stage, we will discuss the effects of multi-granular synthetic queries, synthetic data with domain-specific constraints, and the use of Context2ID data. For the preference learning stage, we will discuss using different candidates for preference learning.
+
+# 5.3.1 Supervised Fine-Tuning Stage
+
+Effects of multi-granular synthetic queries. We conduct an analysis on the effects of incorporating synthetic queries generated from the context at different levels of granularity on MultiHopRAG. We train the generative retrieval model based on semantic identifiers on chunk-level Query2ID data (Chunk), comparing it with the model trained on chunk-level and sentence-level Query2ID data (Chunk+Sent), and both models use Context2ID data. The results are shown in Table 2. We find that sentence-level synthetic queries can significantly improve retrieval performance, indicating that synthetic query generation with a small context can help capture more details from the document.
+
+Effects of constraints-based synthetic queries. We further study the use of constraints-based
+
+<table><tr><td rowspan="2"></td><td colspan="4">MultiHop-RAG</td><td colspan="3">AllSides</td><td colspan="3">AGNews</td></tr><tr><td>HIT@4</td><td>HIT@10</td><td>MAP@10</td><td>MRR@10</td><td>HIT@1</td><td>HIT@5</td><td>HIT@10</td><td>HIT@1</td><td>HIT@5</td><td>HIT@10</td></tr><tr><td>w/o constraints</td><td>61.64</td><td>81.69</td><td>22.13</td><td>47.20</td><td>10.19</td><td>29.63</td><td>47.22</td><td>59.91</td><td>83.94</td><td>88.11</td></tr><tr><td>w/ constraints</td><td>69.98</td><td>88.34</td><td>24.85</td><td>52.29</td><td>14.20</td><td>38.58</td><td>51.85</td><td>62.19</td><td>83.78</td><td>88.24</td></tr></table>
+
+Table 3: Ablation study on generative retrieval performances with or without the constraints-based synthetic queries.  
+
+<table><tr><td rowspan="2"></td><td colspan="4">MultiHop-RAG</td><td colspan="4">Natural Questions</td></tr><tr><td>HIT@4</td><td>HIT@10</td><td>MAP@10</td><td>MRR@10</td><td>HIT@1</td><td>HIT@5</td><td>HIT@10</td><td>MRR@10</td></tr><tr><td>w/o Context2ID</td><td>41.33</td><td>69.31</td><td>14.45</td><td>31.25</td><td>69.72</td><td>85.58</td><td>89.01</td><td>76.57</td></tr><tr><td>w/ Context2ID</td><td>69.98</td><td>88.34</td><td>24.85</td><td>52.29</td><td>70.71</td><td>86.48</td><td>89.85</td><td>77.54</td></tr></table>
+
+Table 4: Ablation study on generative retrieval performance trained with or without Context2ID data. The results demonstrate the helpfulness of Context2ID data and learning to memorize the context for generative retrieval.  
+
+<table><tr><td></td><td>HIT@4</td><td>HIT@10</td><td>MAP@10</td><td>MRR@10</td></tr><tr><td>Concat</td><td>44.30</td><td>72.77</td><td>15.64</td><td>33.59</td></tr><tr><td>Interleave</td><td>69.98</td><td>88.34</td><td>24.85</td><td>52.29</td></tr></table>
+
+Table 5: Analysis on different ways of combining Query2ID and Context2ID data on Multihop-RAG. We compare simple concatenation (Concat) and interleaving (Interleave) that inherently upsamples the Context2ID data.
+
+synthetic queries that are customized for each domain-specific setting. We conduct experiments on three domain-specific corpora, MultiHop-RAG, AllSides, and AGNews. We compare the semantic identifier-based generative retrieval model trained with or without constraints-based synthetic queries, combing with the corresponding Context2ID data. The results are shown in Table 3. The results show that constraints-based synthetic queries can further improve retrieval performance, indicating that it is helpful to use LLM-produced synthetic queries for domain customization.
+
+Effects of Context2ID data. Existing work (Zhuang et al., 2022a; Kishore et al., 2023) debates whether Context2ID data are useful for generative retrieval training. In this work, we consider Context2ID data as an important part of the data recipe, and also include the memorization of the context as part of the supervised fine-tuning objective. Therefore, we conduct an analysis that removes the Context2ID data on MultiHop-RAG and NQ, and the results are shown in Table 4. We can find that Context2ID data consistently improve generative retrieval performance. We also include the comparison of the strategies to combine Query2ID and Context2ID data, including simple concatenation or interleaving Context2ID and
+
+![](images/e0b3c324d9909fd5aa0231eb6cbc6d98a86b4f78f96c69c8b6b797b4a68b1f1b.jpg)  
+Figure 2: Jaccard similarity post-analysis on MultiHopRAG test set. Synthetic queries from Mixtral 8x7b are generally closer to the test set than those from docT5query. Besides, incorporating granularity and domain-specific attributes further helps with getting queries that are closer to the test set.
+
+Query2ID that will upsample Context2ID data (i.e., smaller size dataset) on MultiHop-RAG in Table 5, again illustrating the importance of Context2ID and learning to memorize context may strengthen the effects on Context2ID.
+
+Different query generation models. As we primarily use LLMs to produce synthetic queries, it is important to understand the performance and effects of using an LLM compared to a specialized query generation model. To this end, we conduct a comparison between synthetic queries generated by Mixtral 8x7b and those from docT5query, as shown in Table 6. For a fair comparison, we exclude constraints-based queries because docT5query cannot produce these types of queries. The results show generative retrieval models trained with queries from Mixtral 8x7b consistently outperform models trained on queries from docT5query.
+
+Following Pradeep et al. (2023), we use Jaccard similarity to evaluate the semantic similarity between test queries and synthetic queries as a post-analysis. For each test query, we compute the
+
+<table><tr><td></td><td colspan="4">MultiHop-RAG</td><td colspan="4">Natural Questions</td></tr><tr><td></td><td>HIT@4</td><td>HIT@10</td><td>MAP@10</td><td>MRR@10</td><td>HIT@1</td><td>HIT@5</td><td>HIT@10</td><td>MRR@10</td></tr><tr><td>docT5query</td><td>50.86</td><td>73.30</td><td>17.60</td><td>37.73</td><td>63.30</td><td>79.12</td><td>85.18</td><td>70.30</td></tr><tr><td>Mixtral 8x7b</td><td>61.64</td><td>81.69</td><td>22.13</td><td>47.20</td><td>70.71</td><td>86.48</td><td>89.85</td><td>77.54</td></tr></table>
+
+Table 6: Generative retrieval performance with synthetic queries from Mixtral 8x7b and docT5query. The results show that queries from Mixtral 8x7b can help train a better generative retrieval model.  
+
+<table><tr><td></td><td>HIT@4</td><td>HIT@10</td><td>MAP@10</td><td>MRR@10</td></tr><tr><td>all</td><td>74.32</td><td>88.03</td><td>29.71</td><td>59.26</td></tr><tr><td>w/o Context2ID</td><td>72.15</td><td>86.21</td><td>28.54</td><td>57.50</td></tr><tr><td>w/o Sent</td><td>58.40</td><td>75.17</td><td>21.51</td><td>44.76</td></tr><tr><td>w/o constraints</td><td>68.34</td><td>83.73</td><td>26.28</td><td>53.83</td></tr></table>
+
+Table 7: Ablation study on atomic identifier-based generative retrieval performance on MultiHop-RAG.
+
+maximum Jaccard similarity among all synthetic queries generated for the corresponding chunk or document. We then compute the win rate between two synthetic query sets (e.g., Mixtral 8x7b versus docT5query) as the proportion of test queries for which one set exhibits higher Jaccard similarity. Let  $Q$  be the set of test queries. Then, the win rate for a synthetic query set  $S$  over another set  $\mathcal{T}$  is defined as:
+
+Win  $\operatorname{Rate}(\mathcal{S},\mathcal{T}) =$
+
+$$
+\frac {1}{| Q |} \sum_ {q \in Q} \mathbb {I} \left[ \max _ {s \in \mathcal {S} (q)} J (q, s) > \max _ {t \in \mathcal {T} (q)} J (q, t) \right],
+$$
+
+where  $J(q,s)$  denotes the Jaccard similarity between the token sequences of the test query  $q$  and a synthetic query  $s$ , and  $\mathcal{S}(q)$  represents the subset of synthetic queries for the corresponding chunk.
+
+Figure 2 illustrates that synthetic queries from Mixtral 8x7b generally have a higher similarity to test queries. Moreover, both sentence-level synthetic queries and constraints-based queries contribute to improved distribution matching with the test queries.
+
+Generalization to different identifiers. We further study the generalizability of our data strategies across different types of document identifiers. In this analysis, we use atomic identifier, which are arbitrary unique IDs assigned to each document or chunk. We conduct experiments on MultiHopRAG and the results are shown in Table 7. The findings align with our observations using semantic identifiers, highlighting the critical role of all three data types in generative retrieval. Among them, sentence-level synthetic queries contribute the most to performance improvements.
+
+# 5.3.2 Preference Learning Stage
+
+Effects of negative candidate sources. We first study the strategies of candidates selection for preference learning. We compare random selection from the corpus with using the top candidates from the generative retrieval model after the supervised fine-tuning stage. The results are shown in Table 8, which illustrates that candidate selection has an impact on preference learning, and random candidates may have a negative impact.
+
+Effects of negative candidate number. We also study the effects of using different negative candidate numbers for each query. We experiment with selecting Top-5 and Top-10 negative candidates with a rank higher than the positive candidate from the retrieval results. The results are shown in Table 8. In general, it is effective to use the strategy, which includes high-quality candidates with ranks higher than the corresponding positive candidates. We also see some slight differences when including different numbers of negative candidates. We can find that a large number of negative candidates helps better in metrics such as HIT@1 and HIT@4.
+
+# 5.3.3 Comparison to Off-The-Shelf Retrievers
+
+We also compare the performance of our generative retrieval with several off-the-shelf retrievers, including BM25 (Robertson and Walker, 1994), bge-large-en-v1.5 (Xiao et al., 2023), Contrievermsmarco (Izacard et al., 2021), E5-mistral-7b-instruct (Wang et al., 2024) and GTE-Qwen2-7B-instruct (Li et al., 2023c). The results are shown in Figure 3 and more detailed results can be found in Appendix B.2. The results demonstrate that generative retrieval models, which rely solely on in-domain synthetic data training without retrieval pre-training, can achieve competitive performance compared to those retrievers. These results indicate the potential of generative retrieval, as well as the effectiveness of using LLMs to generate synthetic data tailored to domain-specific needs.
+
+<table><tr><td rowspan="2"></td><td colspan="4">MultiHop-RAG</td><td colspan="4">Natural Questions</td></tr><tr><td>HIT@4</td><td>HIT@10</td><td>MAP@10</td><td>MRR@10</td><td>HIT@1</td><td>HIT@5</td><td>HIT@10</td><td>MRR@10</td></tr><tr><td>SFT</td><td>69.98</td><td>88.34</td><td>24.85</td><td>52.29</td><td>70.71</td><td>86.48</td><td>89.85</td><td>77.54</td></tr><tr><td>Random 5</td><td>58.94</td><td>82.88</td><td>20.88</td><td>43.53</td><td>70.19</td><td>86.48</td><td>89.50</td><td>77.17</td></tr><tr><td>Top-5 negative</td><td>71.53</td><td>89.62</td><td>26.36</td><td>55.40</td><td>71.02</td><td>87.32</td><td>90.04</td><td>78.02</td></tr><tr><td>Top-10 negative</td><td>71.88</td><td>89.80</td><td>26.23</td><td>54.94</td><td>71.22</td><td>87.41</td><td>89.97</td><td>78.14</td></tr></table>
+
+Table 8: Preference learning with different numbers of negative candidates. The results show that it is an effective strategy to select negative candidates with ranks higher than the positive candidate, while different numbers of negative candidates may optimize the retrieval performance in different ways.
+
+![](images/c96f6e8a747f7ee7a9a89b144327fc11eb7f373a02d6ae3c7ca892ee40d83491.jpg)  
+Figure 3: Performance comparison between generative retrieval using semantic identifiers and off-the-shelf-retrieval models. We use HIT@4 for MultiHop-RAG and HIT@1 for other datasets as the evaluation metric. For the full experiment results, please refer to Appendix B.2 (Table 12).
+
+# 6 Related Work
+
+Generative retrieval modeling. Previous work has explored various aspects of generative retrieval. One line of research aims to find appropriate document identifiers for generation, such as numerical or atomic identifier (Tay et al., 2022; Zhuang et al., 2022a; Zhou et al., 2022), N-grams (Bevilacqua et al., 2022; Chen et al., 2023a), titles or URLs (Li et al., 2023a; Lee et al., 2022; Chen et al., 2022; Ziems et al., 2023), keywords-based or summary-based semantic identifiers (Lee et al., 2023b; Tang et al., 2023), codebook (Zhang et al., 2023; Yang et al., 2023; Zeng et al., 2024), and full passages themselves (Tang et al., 2024a). There are also efforts to combine the advantages of different identifiers (Li et al., 2023b). Another line of work tackles the optimization of generative retrieval models, such as incorporating ranking losses (Zhou et al., 2023; Li et al., 2024b; Tang et al., 2024b), or using auxiliary tasks to enhance training (Li et al., 2024a). During retrieval, different constrained decoding methods have been explored to obtain valid identifiers, such as FM-Index (Bevilacqua et al., 2022), Trie-based (Cao et al., 2020), and set-based inference (Tang et al., 2024b).
+
+Synthetic query generation. Alongside the progress in generative retrieval modeling and opti
+
+mization, synthetic query generation has emerged as a pivotal technique for enhancing retrieval systems, particularly in domains with limited annotated data. In dense retrieval, synthetic queries have been used extensively to improve cross-domain performance. For instance, Ma et al. (2021) generated synthetic questions for target-domain documents with a question generation model trained on general-domain data, thereby improving retrieval performance in zero-shot settings. Similarly, Wang et al. (2022) introduced generative pseudo labeling, which combines query generation with pseudo labeling using a cross-encoder to capture finer-grained ranking signals. Further advancements include Bonifacio et al. (2022) and Jeronymo et al. (2023), which leverage large language models to generate synthetic queries in a few-shot manner, and then combine with top K documents ranked by the conditional question generation probability, to train a domain-specific reranker.
+
+Despite these successes in dense retrieval, the potential of synthetic data for generative retrieval has been underexplored. Existing studies typically rely on passage-level synthetic queries generated by docT5query (Nogueira et al., 2019), following the DSI-QG paradigm (Zhuang et al., 2022a). Chen et al. (2023b) explores breaking documents into text fragments for query generation and memo
+
+rization. However, there still lacks a comprehensive discussions on effective strategies for generating synthetic data tailored to domain-specific corpora, especially with LLMs. This work investigates data strategies from multiple perspectives, including the generation of synthetic queries using multi-granularity contexts, incorporating search constraints, and exploring the impact of context data. For preference learning, Zhou et al. (2023) proposes using preference learning objectives for generative retrieval with specialized reward models, though acquiring such models in a domain-specific setting can be challenging. In contrast, our proposed preference learning strategy directly uses the retrieval results to obtain the preference data, offering a more streamlined approach for domain-specific applications.
+
+Similar synthetic-data paradigms have also been explored in Text-to-SQL tasks. Kobayashi et al. (2025) propose YORO, which internalizes database schema and content via LLM-generated NLQ-SQL pairs so that at inference time no schema encoding or value retrieval is required. This parallels our goal of teaching a generative retriever to internalize a domain corpus, and suggests a broader applicability of synthetic data strategies for LLM customization.
+
+# 7 Conclusion
+
+In this work, we explore several strategies to produce synthetic data for generative retrieval training. We find that adding queries in multi-granularity and queries with domain-specific constraints can largely improve the generative retrieval performance during supervised fine-tuning, and memorizing document contents can also contribute to the generative retrieval training. We also find that it is critical to choose high-quality hard negative candidates to effectively use the preference learning objectives to further improve generative retrieval.
+
+# Limitations
+
+Our proposed synthetic data strategies mainly focus on the supervised fine-tuning and preference learning stage. But there are also settings that can largely improve the usability of generative retrieval, such as incremental learning or generalization to unseen documents. It is also important to extend the data strategy exploration for these settings. In addition, similar data strategies may also be effectively used to enhance dense retrieval domain adaptation. Fur
+
+ther systematic research is needed to investigate the strategies for dense retrieval model fine-tuning, as well as the differences between generative and dense model training.
+
+Our synthetic queries are mainly based on one document. However, queries from the real world may be more complex, such as those involving multiple documents with multi-hop reasoning or multi-evidence comparison. It is still under-investigation that to generate those complex queries and to use those queries during retrieval model training.
+
+# References
+
+AI Anthropic. 2024. The claude 3 model family: Opus, sonnet, haiku. *Claude-3 Model Card*, 1.  
+Ramy Baly, Giovanni Da San Martino, James Glass, and Preslav Nakov. 2020. We can detect your bias: Predicting the political ideology of news articles. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4982-4991, Online. Association for Computational Linguistics.  
+Michele Bevilacqua, Giuseppe Ottaviano, Patrick S. H. Lewis, Scott Yih, Sebastian Riedel, and Fabio Petroni. 2022. Autoregressive search engines: Generating substrings as document identifiers. In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9, 2022.  
+Luiz Bonifacio, Hugo Abonizio, Marzieh Fadaee, and Rodrigo Nogueira. 2022. Inpars: Data augmentation for information retrieval using large language models. arXiv preprint arXiv:2202.05144.  
+Nicola De Cao, Gautier Izacard, Sebastian Riedel, and Fabio Petroni. 2020. Autoregressive entity retrieval. CoRR, abs/2010.00904.  
+Jiangui Chen, Ruqing Zhang, Jiafeng Guo, Maarten de Rijke, Yiqun Liu, Yixing Fan, and Xueqi Cheng. 2023a. A unified generative retriever for knowledge-intensive language tasks via prompt learning. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR 2023, Taipei, Taiwan, July 23-27, 2023, pages 1448-1457. ACM.  
+Jiangui Chen, Ruqing Zhang, Jiafeng Guo, Yiqun Liu, Yixing Fan, and Xueqi Cheng. 2022. Corpusbrain: Pre-train a generative retrieval model for knowledge-intensive language tasks. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management, Atlanta, GA, USA, October 17-21, 2022, pages 191-200. ACM.  
+Xiaoyang Chen, Yanjiang Liu, Ben He, Le Sun, and Yingfei Sun. 2023b. Understanding differential
+
+search index for text retrieval. In Findings of the Association for Computational Linguistics: ACL 2023, pages 10701-10717, Toronto, Canada. Association for Computational Linguistics.  
+Tri Dao. 2024. Flashattention-2: Faster attention with better parallelism and work partitioning. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net.  
+Sylvain Gugger, Lysandre Debut, Thomas Wolf, Philipp Schmid, Zachary Mueller, Sourab Mangrulkar, Marc Sun, and Benjamin Bossan. 2022. Accelerate: Training and inference at scale made simple, efficient and adaptable. https://github.com/huggingface/accelerate.  
+Helia Hashemi, Yong Zhuang, Sachith Sri Ram Kothur, Srivas Prasad, Edgar Meij, and W. Bruce Croft. 2023. Dense retrieval adaptation using target domain description. In Proceedings of the 2023 ACM SIGIR International Conference on Theory of Information Retrieval, ICTIR 2023, Taipei, Taiwan, 23 July 2023, pages 95-104. ACM.  
+Gautier Izacard, Mathilde Caron, Lucas Hosseini, Sebastian Riedel, Piotr Bojanowski, Armand Joulin, and Edouard Grave. 2021. Unsupervised dense information retrieval with contrastive learning.  
+Vitor Jeronymo, Luiz Bonifacio, Hugo Abonizio, Marzieh Fadaee, Roberto Lotufo, Jakub Zavrel, and Rodrigo Nogueira. 2023. Inpars-v2: Large language models as efficient dataset generators for information retrieval. arXiv preprint arXiv:2301.01820.  
+Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de Las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, and William El Sayed. 2023. Mistral 7b. CoRR, abs/2310.06825.  
+Albert Q. Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de Las Casas, Emma Bou Hanna, Florian Bressand, Gianna Lengyel, Guillaume Bour, Guillaume Lample, Lélio Renard Lavaud, Lucile Saulnier, Marie-Anne Lachaux, Pierre Stock, Sandeep Subramanian, Sophia Yang, Szymon Antoniak, Teven Le Scao, Théophile Gervet, Thibaut Lavril, Thomas Wang, Timothée Lacroix, and William El Sayed. 2024. Mixtral of experts. CoRR, abs/2401.04088.  
+Vladimir Karpukhin, Barlas Oğuz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen Tau Yih. 2020. Dense passage retrieval for open-domain question answering. In 2020 Conference on Empirical Methods in Natural Language Processing, EMNLP 2020, pages 6769-6781. Association for Computational Linguistics (ACL).
+
+Varsha Kishore, Chao Wan, Justin Lovelace, Yoav Artzi, and Kilian Q. Weinberger. 2023. Incdsci: Incrementally updatable document retrieval. In International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA, volume 202 of Proceedings of Machine Learning Research, pages 17122-17134. PMLR.  
+Hideo Kobayashi, Wuwei Lan, Peng Shi, Shuaichen Chang, Jiang Guo, Henghui Zhu, Zhiguo Wang, and Patrick Ng. 2025. You only read once (YORO): Learning to internalize database knowledge for text-to-SQL. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 1889-1901, Albuquerque, New Mexico. Association for Computational Linguistics.  
+Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, Kristina Toutanova, Llion Jones, Matthew Kelcey, Ming-Wei Chang, Andrew M. Dai, Jakob Uszkoreit, Quoc Le, and Slav Petrov. 2019. Natural questions: A benchmark for question answering research. Transactions of the Association for Computational Linguistics, 7:452-466.  
+Hyunjji Lee, JaeYoung Kim, Hoyeon Chang, Hanseok Oh, Sohee Yang, Vladimir Karpukhin, Yi Lu, and Minjoon Seo. 2023a. Nonparametric decoding for generative retrieval. In Findings of the Association for Computational Linguistics: ACL 2023, pages 12642-12661, Toronto, Canada. Association for Computational Linguistics.  
+Hyunjji Lee, Sohee Yang, Hanseok Oh, and Minjoon Seo. 2022. Generative multi-hop retrieval. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 1417-1436, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.  
+Sunkyung Lee, Minjin Choi, and Jongwuk Lee. 2023b. GLEN: Generative retrieval via lexical index learning. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 7693-7704, Singapore. Association for Computational Linguistics.  
+Xiaoxi Li, Zhicheng Dou, Yujia Zhou, and Fangchao Liu. 2024a. Corpuslm: Towards a unified language model on corpus for knowledge-intensive tasks. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR 2024, Washington DC, USA, July 14-18, 2024, pages 26-37. ACM.  
+Yongqi Li, Nan Yang, Liang Wang, Furu Wei, and Wenjie Li. 2023a. Generative retrieval for conversational question answering. Inf. Process. Manag., 60(5):103475.
+
+Yongqi Li, Nan Yang, Liang Wang, Furu Wei, and Wenjie Li. 2023b. Multiview identifiers enhanced generative retrieval. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6636-6648, Toronto, Canada. Association for Computational Linguistics.  
+Yongqi Li, Nan Yang, Liang Wang, Furu Wei, and Wenjie Li. 2024b. Learning to rank in generative retrieval. In Thirty-Eighth AAAI Conference on Artificial Intelligence, AAAI 2024, Thirty-Sixth Conference on Innovative Applications of Artificial Intelligence, IAAI 2024, Fourteenth Symposium on Educational Advances in Artificial Intelligence, EAAI 2014, February 20-27, 2024, Vancouver, Canada, pages 8716-8723. AAAI Press.  
+Zehan Li, Xin Zhang, Yanzhao Zhang, Dingkun Long, Pengjun Xie, and Meishan Zhang. 2023c. Towards general text embeddings with multi-stage contrastive learning. arXiv preprint arXiv:2308.03281.  
+Ji Ma, Ivan Korotkov, Yinfei Yang, Keith Hall, and Ryan McDonald. 2021. Zero-shot neural passage retrieval via domain-targeted synthetic question generation. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 1075-1088.  
+Rodrigo Nogueira, Jimmy Lin, and AI Epistemic. 2019. From doc2query to docctttttquery.  
+Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul F. Christiano, Jan Leike, and Ryan Lowe. 2022. Training language models to follow instructions with human feedback. In NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9, 2022.  
+Richard Yuanzhe Pang, Weizhe Yuan, Kyunghyun Cho, He He, Sainbayar Sukhbaatar, and Jason Weston. 2024. Iterative reasoning preference optimization. CoRR, abs/2404.19733.  
+Ronak Pradeep, Kai Hui, Jai Gupta, Adam Lelkes, Honglei Zhuang, Jimmy Lin, Donald Metzler, and Vinh Tran. 2023. How does generative retrieval scale to millions of passages? In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 1305-1321, Singapore. Association for Computational Linguistics.  
+Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D. Manning, Stefano Ermon, and Chelsea Finn. 2023. Direct preference optimization: Your language model is secretly a reward model. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023.
+
+Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21:140:1-140:67.  
+Jeff Rasley, Samyam Rajbhandari, Olatunjri Ruwase, and Yuxiong He. 2020. Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters. In KDD '20: The 26th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Virtual Event, CA, USA, August 23-27, 2020, pages 3505-3506. ACM.  
+Stephen E. Robertson and Steve Walker. 1994. Some simple effective approximations to the 2-poisson model for probabilistic weighted retrieval. In Proceedings of the 17th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval. Dublin, Ireland, 3-6 July 1994 (Special Issue of the SIGIR Forum), pages 232-241. ACM/Springer.  
+Weiwei Sun, Lingyong Yan, Zheng Chen, Shuaiqiang Wang, Haichao Zhu, Pengjie Ren, Zhumin Chen, Dawei Yin, Maarten de Rijke, and Zhaochun Ren. 2023. Learning to tokenize for generative retrieval. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023.  
+Qiaoyu Tang, Jiawei Chen, Zhuoqun Li, Bowen Yu, Yaojie Lu, Cheng Fu, Haiyang Yu, Hongyu Lin, Fei Huang, Ben He, Xianpei Han, Le Sun, and Yongbin Li. 2024a. Self-retrieval: End-to-end information retrieval with one large language model. Preprint, arXiv:2403.00801.  
+Yixuan Tang and Yi Yang. 2024. Multihop-rag: Benchmarking retrieval-augmented generation for multi-hop queries. CoRR, abs/2401.15391.  
+Yubao Tang, Ruqing Zhang, Jiafeng Guo, Jiangui Chen, Zuowei Zhu, Shuaiqiang Wang, Dawei Yin, and Xueqi Cheng. 2023. Semantic-enhanced differentiable search index inspired by learning strategies. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD 2023, Long Beach, CA, USA, August 6-10, 2023, pages 4904-4913. ACM.  
+Yubao Tang, Ruqing Zhang, Jiafeng Guo, Maarten de Rijke, Wei Chen, and Xueqi Cheng. 2024b. Listwise generative retrieval models via a sequential learning process. ACM Trans. Inf. Syst., 42(5):133:1-133:31.  
+Yi Tay, Vinh Tran, Mostafa Dehghani, Jianmo Ni, Dara Bahri, Harsh Mehta, Zhen Qin, Kai Hui, Zhe Zhao, Jai Prakash Gupta, Tal Schuster, William W. Cohen, and Donald Metzler. 2022. Transformer memory as a differentiable search index. In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022,
+
+NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9, 2022.  
+Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 5998-6008.  
+Leandro von Werra, Younes Belkada, Lewis Tunstall, Edward Beeching, Tristan Thrush, Nathan Lambert, Shengyi Huang, Kashif Rasul, and Quentin Gallouédec. 2020. Trl: Transformer reinforcement learning. https://github.com/huggingface/trl.  
+Kexin Wang, Nandan Thakur, Nils Reimers, and Iryna Gurevych. 2022. Gpl: Generative pseudo labeling for unsupervised domain adaptation of dense retrieval. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2345-2360.  
+Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. 2024. Improving text embeddings with large language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 11897-11916, Bangkok, Thailand. Association for Computational Linguistics.  
+Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander Rush. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38-45, Online. Association for Computational Linguistics.  
+Shitao Xiao, Zheng Liu, Peitan Zhang, and Niklas Muennighoff. 2023. C-pack: Packaged resources to advance general chinese embedding. Preprint, arXiv:2309.07597.  
+Tianchi Yang, Minghui Song, Zihan Zhang, Haizhen Huang, Weiwei Deng, Feng Sun, and Qi Zhang. 2023. Auto search indexer for end-to-end document retrieval. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 6955-6970, Singapore. Association for Computational Linguistics.  
+Yue Yu, Yuchen Zhuang, Jieyu Zhang, Yu Meng, Alexander J. Ratner, Ranjay Krishna, Jiaming Shen, and Chao Zhang. 2023. Large language model as attributed training data generator: A tale of diversity and bias. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural
+
+Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023.  
+Hansi Zeng, Chen Luo, Bowen Jin, Sheikh Muhammad Sarwar, Tianxin Wei, and Hamed Zamani. 2024. Scalable and effective generative information retrieval. In Proceedings of the ACM on Web Conference 2024, WWW 2024, Singapore, May 13-17, 2024, pages 1441-1452. ACM.  
+Hailin Zhang, Yujing Wang, Qi Chen, Ruiheng Chang, Ting Zhang, Ziming Miao, Yingyan Hou, Yang Ding, Xupeng Miao, Haonan Wang, Bochen Pang, Yuefeng Zhan, Hao Sun, Weiwei Deng, Qi Zhang, Fan Yang, Xing Xie, Mao Yang, and Bin Cui. 2023. Model-enhanced vector index. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023.  
+Peitian Zhang, Zheng Liu, Yujia Zhou, Zhicheng Dou, Fangchao Liu, and Zhao Cao. 2024. Generative retrieval via term set generation. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR 2024, Washington DC, USA, July 14-18, 2024, pages 458-468. ACM.  
+Xinran Zhao, Tong Chen, Sihao Chen, Hongming Zhang, and Tongshuang Wu. 2024. Beyond relevance: Evaluate and improve retrievers on perspective awareness. CoRR, abs/2405.02714.  
+Yujia Zhou, Zhicheng Dou, and Ji-Rong Wen. 2023. Enhancing generative retrieval with reinforcement learning from relevance feedback. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 12481-12490, Singapore. Association for Computational Linguistics.  
+Yujia Zhou, Jing Yao, Zhicheng Dou, Ledell Wu, Peitian Zhang, and Ji-Rong Wen. 2022. Ultron: An ultimate retriever on corpus with a model-based indexer. CoRR, abs/2208.09257.  
+Shengyao Zhuang, Houxing Ren, Linjun Shou, Jian Pei, Ming Gong, Guido Zuccon, and Daxin Jiang. 2022a. Bridging the gap between indexing and retrieval for differentiable search index with query generation. CoRR, abs/2206.10128.  
+Shengyao Zhuang, Houxing Ren, Linjun Shou, Jian Pei, Ming Gong, Guido Zuccon, and Daxin Jiang. 2022b. Bridging the gap between indexing and retrieval for differentiable search index with query generation. arXiv preprint arXiv:2206.10128.  
+Noah Ziems, Wenhao Yu, Zhihan Zhang, and Meng Jiang. 2023. Large language models are built-in autoregressive search engines. In *Findings of the Association for Computational Linguistics: ACL* 2023, pages 2666-2678, Toronto, Canada. Association for Computational Linguistics.
+
+<table><tr><td>Dataset</td><td>Context</td><td>Chunk-Level</td><td>Queries Sentence-Level</td><td>Constraints-Based</td></tr><tr><td>MultiHop-RAG</td><td>7,724</td><td>72,090</td><td>472,193</td><td>51,212</td></tr><tr><td>AllSides</td><td>645</td><td>6,313</td><td>173,898</td><td>6,091</td></tr><tr><td>AGNews</td><td>1,050</td><td>10,355</td><td>80,524</td><td>20,875</td></tr><tr><td>NQ</td><td>98,748</td><td>1,459,031</td><td>-</td><td>-</td></tr></table>
+
+Table 9: Dataset Statistics  
+
+<table><tr><td>Dataset</td><td>Attributes</td></tr><tr><td>MultiHop-RAG</td><td>author, publish time, source, category, title</td></tr><tr><td>AllSides</td><td>political polarity</td></tr><tr><td>AGNews</td><td>location, topic</td></tr></table>
+
+Table 10: Attributes used in each dataset for constraints-based query generation.
+
+# A Details of Experiment Setup
+
+We use Mistral-7B-Instruct-v0.3 as the base model for generative retrieval with the semantic identifier, while use Mistral-7B-v0.3 as the base model for atomic identifier as it is closer to a classification setting.
+
+For supervised fine-tuning, we train the models with 2 epochs, with a learning rate of 2e-5 and a warmup ratio of 0.1. The batch size is set as 256. We use sequence packing to put multiple examples in one forward pass (Raffel et al., 2020). We use bfloat16 for our training.
+
+For preference learning, we mainly conduct experiments on MultiHop-RAG and NQ with semantic identifiers. We train the models with 1 epoch. The learning rate is set as 1e-7, batch size is set as 64,  $\beta$  is set as  $0.5$ ,  $\alpha$  is set as 1.0.
+
+The training infrastructure includes TRL (von Werra et al., 2020), Accelerate (Gugger et al., 2022), Transformers (Wolf et al., 2020), DeepSpeed (Rasley et al., 2020) and FlashAttention-2 (Dao, 2024). We use 8x Nvidia A100-SXM4-40GB for our experiments. Each training or inference procedure can be completed in 1 day.
+
+Statistics of the numbers of the documents, different synthetic queries can be found in Table 9. Attributes used for constraints-based synthetic queries can be found in Table 10.
+
+# A.1 MultiHop-RAG
+
+On MultiHop-RAG, we split the documents into chunks with maximum length of 256 without overlap and conduct retrieval on individual chunks. For synthetic query generation,  $m_c$ ,  $m_s$  and  $m_i$  are set
+
+as 10, and the temperature for LLM inference on synthetic data generation is set as 0.7. We interleave the Context2ID and Query2ID data as the full dataset for model supervised fine-tuning. The maximum sequence length is set as 700. For synthetic queries for preference learning, we ask the LLM to generate 10 queries. We perform the retrieval with beam size as 10 and retrieve the top-10 candidates for each query to construct the candidate pairs.
+
+# A.2 AllSides
+
+On AllSides, we conduct document-level retrieval. For synthetic query generation,  $m_c$ ,  $m_s$  and  $m_i$  are set as 10, and the temperature for LLM inference on synthetic data generation is set as 0.7. For Context2ID data, as there are some long documents in the corpus, we will split the long context into chunks with maximum length of 256 without overlap. The Context2ID data is constructed to use all chunks in the document to predict its corresponding document identifier. We interleave the Context2ID and Query2ID data as the full dataset for model supervised fine-tuning. The maximum sequence length is set as 700.
+
+# A.3 AGNews
+
+On AGNews, we conduct document-level retrieval. For synthetic query generation,  $m_c$ ,  $m_s$  and  $m_i$  are set as 10, and the temperature for LLM inference on synthetic data generation is set as 0.7. Queries constructed by Zhao et al. (2024) use two different perspectives. The first perspective is either the location of the desired news or the topic, while the second perspective is that the news is similar to another given news in the query. As we mentioned Section 5.2, we replace the second perspective with the other field so that each query consists of both location and topic perspectives. The topic and location information used for instruction-based synthetic query generation is extracted with Mixtral 8x7b. We interleave the Context2ID and Query2ID data as the full dataset for supervised fine-tuning. The maximum sequence length is set as 700.
+
+# A.4 Natural Questions
+
+On NQ, we conduct document-level retrieval. We use the document prefixes from (Kishore et al., 2023) to produce the semantic identifiers. For synthetic query generation, we perform truncation on pages when they are too long so that we always have at least 1024 token space for model output. We set  $m_c$  as 15 and temperature as 0.7. We do not include sentence-level synthetic queries as the number of those queries are too large to be included in training within a reasonable time. Instead, we include sentence-level Context2ID as the approximation, and use the sentences from the document prefixes from (Kishore et al., 2023) to predict corresponding document identifiers. In NQ, we have high quality human annotated training queries, which we also include as part of the Query2ID data and therefore we do not include instruction-based synthetic queries. We concatenate the Context2ID and Query2ID data as the full dataset for model supervised fine-tuning, as interleaving will produce a much larger dataset that cannot be trained within a reasonable time. The maximum sequence length is set as 450. For synthetic queries for preference learning, we also perform truncation as for supervised fine-tuning, and ask the LLM to generate 10 queries. As the generated query number is quite large for inference, we use the first 2 generated queries for each documents for preference learning. We perform the retrieval with beam size as 10 and retrieve the top-10 candidates for each query to construct the candidate pairs.
+
+# B Additional Comparisons
+
+# B.1 Results on Comparison with Filtered Queries
+
+We conduct another experiment to more closely follow the data pipeline of DSI-QG (Zhuang et al., 2022b), where we generate a set of raw queries using each chunk first and then use the BGE reranker (Xiao et al., 2023) to find the top relevant synthetic queries with docT5query on MultiHopRAG. DSI-QG is originally trained with T5 models. For a fair comparison, we still train the same Mistral 7b model, compared to using Mixtral  $8 \times 7$  as the synthetic data generator for chunk-level synthetic queries. The results are shown in Table 11. We can find that our proposed synthetic query strategy still largely outperforms the filtering-based strategy (Chunk-F).
+
+<table><tr><td></td><td>HIT@4</td><td>HIT@10</td><td>MAP@10</td><td>MRR@10</td></tr><tr><td>Chunk-F</td><td>30.69</td><td>59.78</td><td>10.41</td><td>23.27</td></tr><tr><td>Ours (Chunk)</td><td>43.64</td><td>66.65</td><td>13.98</td><td>31.14</td></tr><tr><td>Ours (Full)</td><td>69.98</td><td>88.34</td><td>24.85</td><td>52.29</td></tr></table>
+
+Table 11: Comparison between the filtered queries that follow DSI-QG data pipeline and our synthetic data generation strategy.
+
+# B.2 Results on Comparison with Off-The-Shelf Retrieval Models
+
+The detailed results for each dataset are shown in Table 12. We run the retrieval models on MultiHopRAG, NQ and AGNews to collect the results, and adopt the results of AllSides from Zhao et al. (2024).
+
+# C LLM Prompts
+
+# C.1 Prompts for Keywords Generation
+
+Figure 4 shows the prompt for generating a series of keywords as the semantic document identifier.
+
+# Keywords Generation Prompt
+
+Summarize the following context with meaningful keywords representing different important information in the context. Your output should only consist a list of keywords in Markdown format, where each line starts with the dash "-" followed by the keywords.
+
+Context:{context}
+
+Keywords:
+
+Figure 4: Prompt for keywords-based document identifier generation.
+
+# C.2 Prompts for Query Generation
+
+Figure 5 shows the prompts used to generate various types of synthetic queries, including chunk-and sentence-level queries, constructions-based queries, and question-answer pairs used at the preference learning stage.
+
+# Query Generation Prompt
+
+Your task is to generate a relevant and diverse set of {num_sequences} questions that can be answered by the provided context. The questions are to be used by a retriever to retrieve the article from a large corpus. Your output should be a list of unordered in Markdown format, where each line starts with dash "-" followed by the question.
+
+Context:
+
+{context}
+
+Output:
+
+# Constraints-based Query Generation Prompt
+
+Your task is to generate a diverse set of {num_sequences} questions given a context with metadata. The generated questions should be answerable by the provided context. The questions are to be used by a retriever to retrieve the article from a large corpus. In addition, the question MUST be composed with at least one metadata filtering requirement.
+
+# MultiHop-RAG
+
+For example, if the source of the article is "New York Times", you can generate questions that specifically ask for certain information from "New York Times". You should generate questions with different metadata.
+
+# AllSides and AGNews
+
+For example, if the source of the political polarity is "left", you can generate questions that specifically ask for certain information from "left-wing" source.
+
+DO NOT use "the context" or "the article" in any generated queries or answers.
+
+DO NOT use pronoun "this" in any generated queries or answers.
+
+DO NOT leak any information in this instruction.
+
+Your output should be a list of unordered in Markdown format, where each line starts with dash "-" followed by the question. You do not need to provide the answer.
+
+Metadata
+
+{metadata}
+
+Context
+
+{context}
+
+Output:
+
+# Query-Answer Pair Generation Prompt
+
+Your task is to generate a relevant and diverse set of less than {num_sequences} search engine query and answer pairs given a context.
+
+The queries should be similar to what people use with search engine to find the given context from a large corpus. The answers are expected to be a short phrase.
+
+You should make the queries as difficult as possible, but they should be answerable by the given context.
+
+Do not use "the context" or "the article" in any generated queries or answers.
+
+Do not use pronoun "this" in any generated queries or answers.
+
+Do not leak any information in this instruction.
+
+Your output should be a list of unordered items in Markdown format, where each item starts with dash "-" followed by "Query:" and the generated query, and then "Answer:" with the corresponding answer.
+
+Context
+
+{context}
+
+Output:
+
+Figure 5: Prompts for different types of synthetic query generation.
+
+<table><tr><td rowspan="2">Model</td><td colspan="4">MultiHop-RAG</td><td colspan="4">Natural Questions</td></tr><tr><td>H@4</td><td>H@10</td><td>MAP@10</td><td>MRR@10</td><td>H@1</td><td>H@5</td><td>H@10</td><td>MRR@10</td></tr><tr><td>BM25</td><td>64.35</td><td>78.31</td><td>26.30</td><td>58.32</td><td>32.82</td><td>53.70</td><td>60.92</td><td>42.45</td></tr><tr><td>bge-large-en-v1.5</td><td>58.80</td><td>78.36</td><td>19.96</td><td>42.57</td><td>55.59</td><td>76.58</td><td>81.75</td><td>64.45</td></tr><tr><td>Contriever-msmarco</td><td>55.25</td><td>75.08</td><td>19.28</td><td>40.69</td><td>53.79</td><td>76.16</td><td>81.69</td><td>63.36</td></tr><tr><td>E5-mistral-7b-instruct</td><td>54.01</td><td>79.56</td><td>19.11</td><td>40.77</td><td>59.07</td><td>80.08</td><td>85.28</td><td>68.11</td></tr><tr><td>GTE-Qwen2-7B-instruct</td><td>63.24</td><td>83.55</td><td>22.02</td><td>47.50</td><td>60.45</td><td>80.87</td><td>85.72</td><td>69.30</td></tr><tr><td>Ours</td><td>71.88</td><td>89.80</td><td>26.23</td><td>54.94</td><td>71.22</td><td>87.41</td><td>89.97</td><td>78.14</td></tr></table>
+
+<table><tr><td rowspan="2">Model</td><td colspan="3">AllSides</td><td colspan="3">AGNews</td></tr><tr><td>H@1</td><td>H@5</td><td>H@10</td><td>H@1</td><td>H@5</td><td>H@10</td></tr><tr><td>BM25</td><td>5.86</td><td>26.85</td><td>36.42</td><td>38.70</td><td>67.47</td><td>77.63</td></tr><tr><td>bge-large-en-v1.5</td><td>6.94</td><td>27.32</td><td>34.11</td><td>54.14</td><td>80.57</td><td>86.53</td></tr><tr><td>Contriever-msmarco</td><td>6.64</td><td>25.77</td><td>38.43</td><td>52.69</td><td>80.40</td><td>85.79</td></tr><tr><td>E5-mistral-7b-instruct</td><td>8.18</td><td>28.24</td><td>39.82</td><td>57.32</td><td>85.90</td><td>88.98</td></tr><tr><td>GTE-Qwen2-7B-instruct</td><td>9.11</td><td>34.11</td><td>49.07</td><td>57.65</td><td>83.37</td><td>88.57</td></tr><tr><td>Ours</td><td>14.20</td><td>38.58</td><td>51.85</td><td>62.19</td><td>83.78</td><td>88.24</td></tr></table>
+
+Table 12: Comparisons to Off-The-Shelf retrieval models across four benchmarks.

ACL/2025/On Synthetic Data Strategies for Domain-Specific Generative Retrieval/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:652b53ec0d1ae139ca4cb5574ccf14bc4ddb2d6e2a5bea396308e2e3ab176d0a
+size 523592

ACL/2025/On Synthetic Data Strategies for Domain-Specific Generative Retrieval/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:81e33b8e867d5c3e9fafb35a1bfffd50b469887ab475daabc69cae03cec9a16b
+size 463161

ACL/2025/On the Acquisition of Shared Grammatical Representations in Bilingual Language Models/ef87d9e3-8343-44dd-9167-d771a1a1e32b_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:81727bb85094482f67e2658c66f8e58454aec5ee3dccc59f6ea0ad75e6fedeff
+size 123085

ACL/2025/On the Acquisition of Shared Grammatical Representations in Bilingual Language Models/ef87d9e3-8343-44dd-9167-d771a1a1e32b_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a5c5910fe7cfe21902c5ea5d8c2cab9edc2879343e2c472b5f982c71196a1ee1
+size 149500